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/sched/mm.h>
50 #include <linux/sched/task.h>
51 #include <linux/pagemap.h>
52 #include <linux/swap.h>
53 #include <linux/swapops.h>
54 #include <linux/slab.h>
55 #include <linux/init.h>
56 #include <linux/ksm.h>
57 #include <linux/rmap.h>
58 #include <linux/rcupdate.h>
59 #include <linux/export.h>
60 #include <linux/memcontrol.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/migrate.h>
63 #include <linux/hugetlb.h>
64 #include <linux/backing-dev.h>
65 #include <linux/page_idle.h>
66 #include <linux/memremap.h>
67 #include <linux/userfaultfd_k.h>
69 #include <asm/tlbflush.h>
71 #include <trace/events/tlb.h>
75 static struct kmem_cache *anon_vma_cachep;
76 static struct kmem_cache *anon_vma_chain_cachep;
78 static inline struct anon_vma *anon_vma_alloc(void)
80 struct anon_vma *anon_vma;
82 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
84 atomic_set(&anon_vma->refcount, 1);
85 anon_vma->num_children = 0;
86 anon_vma->num_active_vmas = 0;
87 anon_vma->parent = anon_vma;
89 * Initialise the anon_vma root to point to itself. If called
90 * from fork, the root will be reset to the parents anon_vma.
92 anon_vma->root = anon_vma;
98 static inline void anon_vma_free(struct anon_vma *anon_vma)
100 VM_BUG_ON(atomic_read(&anon_vma->refcount));
103 * Synchronize against page_lock_anon_vma_read() such that
104 * we can safely hold the lock without the anon_vma getting
107 * Relies on the full mb implied by the atomic_dec_and_test() from
108 * put_anon_vma() against the acquire barrier implied by
109 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
111 * page_lock_anon_vma_read() VS put_anon_vma()
112 * down_read_trylock() atomic_dec_and_test()
114 * atomic_read() rwsem_is_locked()
116 * LOCK should suffice since the actual taking of the lock must
117 * happen _before_ what follows.
120 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
121 anon_vma_lock_write(anon_vma);
122 anon_vma_unlock_write(anon_vma);
125 kmem_cache_free(anon_vma_cachep, anon_vma);
128 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
130 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
133 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
135 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
138 static void anon_vma_chain_link(struct vm_area_struct *vma,
139 struct anon_vma_chain *avc,
140 struct anon_vma *anon_vma)
143 avc->anon_vma = anon_vma;
144 list_add(&avc->same_vma, &vma->anon_vma_chain);
145 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
149 * __anon_vma_prepare - attach an anon_vma to a memory region
150 * @vma: the memory region in question
152 * This makes sure the memory mapping described by 'vma' has
153 * an 'anon_vma' attached to it, so that we can associate the
154 * anonymous pages mapped into it with that anon_vma.
156 * The common case will be that we already have one, which
157 * is handled inline by anon_vma_prepare(). But if
158 * not we either need to find an adjacent mapping that we
159 * can re-use the anon_vma from (very common when the only
160 * reason for splitting a vma has been mprotect()), or we
161 * allocate a new one.
163 * Anon-vma allocations are very subtle, because we may have
164 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
165 * and that may actually touch the spinlock even in the newly
166 * allocated vma (it depends on RCU to make sure that the
167 * anon_vma isn't actually destroyed).
169 * As a result, we need to do proper anon_vma locking even
170 * for the new allocation. At the same time, we do not want
171 * to do any locking for the common case of already having
174 * This must be called with the mmap_sem held for reading.
176 int __anon_vma_prepare(struct vm_area_struct *vma)
178 struct mm_struct *mm = vma->vm_mm;
179 struct anon_vma *anon_vma, *allocated;
180 struct anon_vma_chain *avc;
184 avc = anon_vma_chain_alloc(GFP_KERNEL);
188 anon_vma = find_mergeable_anon_vma(vma);
191 anon_vma = anon_vma_alloc();
192 if (unlikely(!anon_vma))
193 goto out_enomem_free_avc;
194 anon_vma->num_children++; /* self-parent link for new root */
195 allocated = anon_vma;
198 anon_vma_lock_write(anon_vma);
199 /* page_table_lock to protect against threads */
200 spin_lock(&mm->page_table_lock);
201 if (likely(!vma->anon_vma)) {
202 vma->anon_vma = anon_vma;
203 anon_vma_chain_link(vma, avc, anon_vma);
204 anon_vma->num_active_vmas++;
208 spin_unlock(&mm->page_table_lock);
209 anon_vma_unlock_write(anon_vma);
211 if (unlikely(allocated))
212 put_anon_vma(allocated);
214 anon_vma_chain_free(avc);
219 anon_vma_chain_free(avc);
225 * This is a useful helper function for locking the anon_vma root as
226 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
229 * Such anon_vma's should have the same root, so you'd expect to see
230 * just a single mutex_lock for the whole traversal.
232 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
234 struct anon_vma *new_root = anon_vma->root;
235 if (new_root != root) {
236 if (WARN_ON_ONCE(root))
237 up_write(&root->rwsem);
239 down_write(&root->rwsem);
244 static inline void unlock_anon_vma_root(struct anon_vma *root)
247 up_write(&root->rwsem);
251 * Attach the anon_vmas from src to dst.
252 * Returns 0 on success, -ENOMEM on failure.
254 * If dst->anon_vma is NULL this function tries to find and reuse existing
255 * anon_vma which has no vmas and only one child anon_vma. This prevents
256 * degradation of anon_vma hierarchy to endless linear chain in case of
257 * constantly forking task. On the other hand, an anon_vma with more than one
258 * child isn't reused even if there was no alive vma, thus rmap walker has a
259 * good chance of avoiding scanning the whole hierarchy when it searches where
262 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
264 struct anon_vma_chain *avc, *pavc;
265 struct anon_vma *root = NULL;
267 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
268 struct anon_vma *anon_vma;
270 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
271 if (unlikely(!avc)) {
272 unlock_anon_vma_root(root);
274 avc = anon_vma_chain_alloc(GFP_KERNEL);
278 anon_vma = pavc->anon_vma;
279 root = lock_anon_vma_root(root, anon_vma);
280 anon_vma_chain_link(dst, avc, anon_vma);
283 * Reuse existing anon_vma if it has no vma and only one
286 * Root anon_vma is never reused:
287 * it has self-parent reference and at least one child.
289 if (!dst->anon_vma &&
290 anon_vma->num_children < 2 &&
291 anon_vma->num_active_vmas == 0)
292 dst->anon_vma = anon_vma;
295 dst->anon_vma->num_active_vmas++;
296 unlock_anon_vma_root(root);
301 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
302 * decremented in unlink_anon_vmas().
303 * We can safely do this because callers of anon_vma_clone() don't care
304 * about dst->anon_vma if anon_vma_clone() failed.
306 dst->anon_vma = NULL;
307 unlink_anon_vmas(dst);
312 * Attach vma to its own anon_vma, as well as to the anon_vmas that
313 * the corresponding VMA in the parent process is attached to.
314 * Returns 0 on success, non-zero on failure.
316 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
318 struct anon_vma_chain *avc;
319 struct anon_vma *anon_vma;
322 /* Don't bother if the parent process has no anon_vma here. */
326 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
327 vma->anon_vma = NULL;
330 * First, attach the new VMA to the parent VMA's anon_vmas,
331 * so rmap can find non-COWed pages in child processes.
333 error = anon_vma_clone(vma, pvma);
337 /* An existing anon_vma has been reused, all done then. */
341 /* Then add our own anon_vma. */
342 anon_vma = anon_vma_alloc();
345 anon_vma->num_active_vmas++;
346 avc = anon_vma_chain_alloc(GFP_KERNEL);
348 goto out_error_free_anon_vma;
351 * The root anon_vma's spinlock is the lock actually used when we
352 * lock any of the anon_vmas in this anon_vma tree.
354 anon_vma->root = pvma->anon_vma->root;
355 anon_vma->parent = pvma->anon_vma;
357 * With refcounts, an anon_vma can stay around longer than the
358 * process it belongs to. The root anon_vma needs to be pinned until
359 * this anon_vma is freed, because the lock lives in the root.
361 get_anon_vma(anon_vma->root);
362 /* Mark this anon_vma as the one where our new (COWed) pages go. */
363 vma->anon_vma = anon_vma;
364 anon_vma_lock_write(anon_vma);
365 anon_vma_chain_link(vma, avc, anon_vma);
366 anon_vma->parent->num_children++;
367 anon_vma_unlock_write(anon_vma);
371 out_error_free_anon_vma:
372 put_anon_vma(anon_vma);
374 unlink_anon_vmas(vma);
378 void unlink_anon_vmas(struct vm_area_struct *vma)
380 struct anon_vma_chain *avc, *next;
381 struct anon_vma *root = NULL;
384 * Unlink each anon_vma chained to the VMA. This list is ordered
385 * from newest to oldest, ensuring the root anon_vma gets freed last.
387 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
388 struct anon_vma *anon_vma = avc->anon_vma;
390 root = lock_anon_vma_root(root, anon_vma);
391 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
394 * Leave empty anon_vmas on the list - we'll need
395 * to free them outside the lock.
397 if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
398 anon_vma->parent->num_children--;
402 list_del(&avc->same_vma);
403 anon_vma_chain_free(avc);
406 vma->anon_vma->num_active_vmas--;
407 unlock_anon_vma_root(root);
410 * Iterate the list once more, it now only contains empty and unlinked
411 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
412 * needing to write-acquire the anon_vma->root->rwsem.
414 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
415 struct anon_vma *anon_vma = avc->anon_vma;
417 VM_WARN_ON(anon_vma->num_children);
418 VM_WARN_ON(anon_vma->num_active_vmas);
419 put_anon_vma(anon_vma);
421 list_del(&avc->same_vma);
422 anon_vma_chain_free(avc);
426 static void anon_vma_ctor(void *data)
428 struct anon_vma *anon_vma = data;
430 init_rwsem(&anon_vma->rwsem);
431 atomic_set(&anon_vma->refcount, 0);
432 anon_vma->rb_root = RB_ROOT_CACHED;
435 void __init anon_vma_init(void)
437 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
438 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
440 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
441 SLAB_PANIC|SLAB_ACCOUNT);
445 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
447 * Since there is no serialization what so ever against page_remove_rmap()
448 * the best this function can do is return a locked anon_vma that might
449 * have been relevant to this page.
451 * The page might have been remapped to a different anon_vma or the anon_vma
452 * returned may already be freed (and even reused).
454 * In case it was remapped to a different anon_vma, the new anon_vma will be a
455 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
456 * ensure that any anon_vma obtained from the page will still be valid for as
457 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
459 * All users of this function must be very careful when walking the anon_vma
460 * chain and verify that the page in question is indeed mapped in it
461 * [ something equivalent to page_mapped_in_vma() ].
463 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
464 * that the anon_vma pointer from page->mapping is valid if there is a
465 * mapcount, we can dereference the anon_vma after observing those.
467 struct anon_vma *page_get_anon_vma(struct page *page)
469 struct anon_vma *anon_vma = NULL;
470 unsigned long anon_mapping;
473 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
474 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
476 if (!page_mapped(page))
479 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
480 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
486 * If this page is still mapped, then its anon_vma cannot have been
487 * freed. But if it has been unmapped, we have no security against the
488 * anon_vma structure being freed and reused (for another anon_vma:
489 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
490 * above cannot corrupt).
492 if (!page_mapped(page)) {
494 put_anon_vma(anon_vma);
504 * Similar to page_get_anon_vma() except it locks the anon_vma.
506 * Its a little more complex as it tries to keep the fast path to a single
507 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
508 * reference like with page_get_anon_vma() and then block on the mutex.
510 struct anon_vma *page_lock_anon_vma_read(struct page *page)
512 struct anon_vma *anon_vma = NULL;
513 struct anon_vma *root_anon_vma;
514 unsigned long anon_mapping;
517 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
518 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
520 if (!page_mapped(page))
523 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
524 root_anon_vma = READ_ONCE(anon_vma->root);
525 if (down_read_trylock(&root_anon_vma->rwsem)) {
527 * If the page is still mapped, then this anon_vma is still
528 * its anon_vma, and holding the mutex ensures that it will
529 * not go away, see anon_vma_free().
531 if (!page_mapped(page)) {
532 up_read(&root_anon_vma->rwsem);
538 /* trylock failed, we got to sleep */
539 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
544 if (!page_mapped(page)) {
546 put_anon_vma(anon_vma);
550 /* we pinned the anon_vma, its safe to sleep */
552 anon_vma_lock_read(anon_vma);
554 if (atomic_dec_and_test(&anon_vma->refcount)) {
556 * Oops, we held the last refcount, release the lock
557 * and bail -- can't simply use put_anon_vma() because
558 * we'll deadlock on the anon_vma_lock_write() recursion.
560 anon_vma_unlock_read(anon_vma);
561 __put_anon_vma(anon_vma);
572 void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
574 anon_vma_unlock_read(anon_vma);
577 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
579 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
580 * important if a PTE was dirty when it was unmapped that it's flushed
581 * before any IO is initiated on the page to prevent lost writes. Similarly,
582 * it must be flushed before freeing to prevent data leakage.
584 void try_to_unmap_flush(void)
586 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
588 if (!tlb_ubc->flush_required)
591 arch_tlbbatch_flush(&tlb_ubc->arch);
592 tlb_ubc->flush_required = false;
593 tlb_ubc->writable = false;
596 /* Flush iff there are potentially writable TLB entries that can race with IO */
597 void try_to_unmap_flush_dirty(void)
599 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
601 if (tlb_ubc->writable)
602 try_to_unmap_flush();
605 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
607 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
609 arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
610 tlb_ubc->flush_required = true;
613 * Ensure compiler does not re-order the setting of tlb_flush_batched
614 * before the PTE is cleared.
617 mm->tlb_flush_batched = true;
620 * If the PTE was dirty then it's best to assume it's writable. The
621 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
622 * before the page is queued for IO.
625 tlb_ubc->writable = true;
629 * Returns true if the TLB flush should be deferred to the end of a batch of
630 * unmap operations to reduce IPIs.
632 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
634 bool should_defer = false;
636 if (!(flags & TTU_BATCH_FLUSH))
639 /* If remote CPUs need to be flushed then defer batch the flush */
640 if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
648 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
649 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
650 * operation such as mprotect or munmap to race between reclaim unmapping
651 * the page and flushing the page. If this race occurs, it potentially allows
652 * access to data via a stale TLB entry. Tracking all mm's that have TLB
653 * batching in flight would be expensive during reclaim so instead track
654 * whether TLB batching occurred in the past and if so then do a flush here
655 * if required. This will cost one additional flush per reclaim cycle paid
656 * by the first operation at risk such as mprotect and mumap.
658 * This must be called under the PTL so that an access to tlb_flush_batched
659 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
662 void flush_tlb_batched_pending(struct mm_struct *mm)
664 if (mm->tlb_flush_batched) {
668 * Do not allow the compiler to re-order the clearing of
669 * tlb_flush_batched before the tlb is flushed.
672 mm->tlb_flush_batched = false;
676 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
680 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
684 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
687 * At what user virtual address is page expected in vma?
688 * Caller should check the page is actually part of the vma.
690 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
692 if (PageAnon(page)) {
693 struct anon_vma *page__anon_vma = page_anon_vma(page);
695 * Note: swapoff's unuse_vma() is more efficient with this
696 * check, and needs it to match anon_vma when KSM is active.
698 if (!vma->anon_vma || !page__anon_vma ||
699 vma->anon_vma->root != page__anon_vma->root)
701 } else if (!vma->vm_file) {
703 } else if (vma->vm_file->f_mapping != compound_head(page)->mapping) {
707 return vma_address(page, vma);
710 pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
718 pgd = pgd_offset(mm, address);
719 if (!pgd_present(*pgd))
722 p4d = p4d_offset(pgd, address);
723 if (!p4d_present(*p4d))
726 pud = pud_offset(p4d, address);
727 if (!pud_present(*pud))
730 pmd = pmd_offset(pud, address);
732 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
733 * without holding anon_vma lock for write. So when looking for a
734 * genuine pmde (in which to find pte), test present and !THP together.
738 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
744 struct page_referenced_arg {
747 unsigned long vm_flags;
748 struct mem_cgroup *memcg;
751 * arg: page_referenced_arg will be passed
753 static bool page_referenced_one(struct page *page, struct vm_area_struct *vma,
754 unsigned long address, void *arg)
756 struct page_referenced_arg *pra = arg;
757 struct page_vma_mapped_walk pvmw = {
764 while (page_vma_mapped_walk(&pvmw)) {
765 address = pvmw.address;
767 if (vma->vm_flags & VM_LOCKED) {
768 page_vma_mapped_walk_done(&pvmw);
769 pra->vm_flags |= VM_LOCKED;
770 return false; /* To break the loop */
774 if (ptep_clear_flush_young_notify(vma, address,
777 * Don't treat a reference through
778 * a sequentially read mapping as such.
779 * If the page has been used in another mapping,
780 * we will catch it; if this other mapping is
781 * already gone, the unmap path will have set
782 * PG_referenced or activated the page.
784 if (likely(!(vma->vm_flags & VM_SEQ_READ)))
787 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
788 if (pmdp_clear_flush_young_notify(vma, address,
792 /* unexpected pmd-mapped page? */
800 clear_page_idle(page);
801 if (test_and_clear_page_young(page))
806 pra->vm_flags |= vma->vm_flags;
810 return false; /* To break the loop */
815 static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
817 struct page_referenced_arg *pra = arg;
818 struct mem_cgroup *memcg = pra->memcg;
820 if (!mm_match_cgroup(vma->vm_mm, memcg))
827 * page_referenced - test if the page was referenced
828 * @page: the page to test
829 * @is_locked: caller holds lock on the page
830 * @memcg: target memory cgroup
831 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
833 * Quick test_and_clear_referenced for all mappings to a page,
834 * returns the number of ptes which referenced the page.
836 int page_referenced(struct page *page,
838 struct mem_cgroup *memcg,
839 unsigned long *vm_flags)
842 struct page_referenced_arg pra = {
843 .mapcount = total_mapcount(page),
846 struct rmap_walk_control rwc = {
847 .rmap_one = page_referenced_one,
849 .anon_lock = page_lock_anon_vma_read,
853 if (!page_mapped(page))
856 if (!page_rmapping(page))
859 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
860 we_locked = trylock_page(page);
866 * If we are reclaiming on behalf of a cgroup, skip
867 * counting on behalf of references from different
871 rwc.invalid_vma = invalid_page_referenced_vma;
874 rmap_walk(page, &rwc);
875 *vm_flags = pra.vm_flags;
880 return pra.referenced;
883 static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma,
884 unsigned long address, void *arg)
886 struct page_vma_mapped_walk pvmw = {
892 unsigned long start = address, end;
896 * We have to assume the worse case ie pmd for invalidation. Note that
897 * the page can not be free from this function.
899 end = vma_address_end(page, vma);
900 mmu_notifier_invalidate_range_start(vma->vm_mm, start, end);
902 while (page_vma_mapped_walk(&pvmw)) {
903 unsigned long cstart, cend;
906 cstart = address = pvmw.address;
909 pte_t *pte = pvmw.pte;
911 if (!pte_dirty(*pte) && !pte_write(*pte))
914 flush_cache_page(vma, address, pte_pfn(*pte));
915 entry = ptep_clear_flush(vma, address, pte);
916 entry = pte_wrprotect(entry);
917 entry = pte_mkclean(entry);
918 set_pte_at(vma->vm_mm, address, pte, entry);
919 cend = cstart + PAGE_SIZE;
922 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
923 pmd_t *pmd = pvmw.pmd;
926 if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
929 flush_cache_page(vma, address, page_to_pfn(page));
930 entry = pmdp_huge_clear_flush(vma, address, pmd);
931 entry = pmd_wrprotect(entry);
932 entry = pmd_mkclean(entry);
933 set_pmd_at(vma->vm_mm, address, pmd, entry);
935 cend = cstart + PMD_SIZE;
938 /* unexpected pmd-mapped page? */
944 mmu_notifier_invalidate_range(vma->vm_mm, cstart, cend);
949 mmu_notifier_invalidate_range_end(vma->vm_mm, start, end);
954 static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
956 if (vma->vm_flags & VM_SHARED)
962 int page_mkclean(struct page *page)
965 struct address_space *mapping;
966 struct rmap_walk_control rwc = {
967 .arg = (void *)&cleaned,
968 .rmap_one = page_mkclean_one,
969 .invalid_vma = invalid_mkclean_vma,
972 BUG_ON(!PageLocked(page));
974 if (!page_mapped(page))
977 mapping = page_mapping(page);
981 rmap_walk(page, &rwc);
985 EXPORT_SYMBOL_GPL(page_mkclean);
988 * page_move_anon_rmap - move a page to our anon_vma
989 * @page: the page to move to our anon_vma
990 * @vma: the vma the page belongs to
992 * When a page belongs exclusively to one process after a COW event,
993 * that page can be moved into the anon_vma that belongs to just that
994 * process, so the rmap code will not search the parent or sibling
997 void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
999 struct anon_vma *anon_vma = vma->anon_vma;
1001 page = compound_head(page);
1003 VM_BUG_ON_PAGE(!PageLocked(page), page);
1004 VM_BUG_ON_VMA(!anon_vma, vma);
1006 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1008 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1009 * simultaneously, so a concurrent reader (eg page_referenced()'s
1010 * PageAnon()) will not see one without the other.
1012 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1016 * __page_set_anon_rmap - set up new anonymous rmap
1017 * @page: Page to add to rmap
1018 * @vma: VM area to add page to.
1019 * @address: User virtual address of the mapping
1020 * @exclusive: the page is exclusively owned by the current process
1022 static void __page_set_anon_rmap(struct page *page,
1023 struct vm_area_struct *vma, unsigned long address, int exclusive)
1025 struct anon_vma *anon_vma = vma->anon_vma;
1033 * If the page isn't exclusively mapped into this vma,
1034 * we must use the _oldest_ possible anon_vma for the
1038 anon_vma = anon_vma->root;
1040 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1041 page->mapping = (struct address_space *) anon_vma;
1042 page->index = linear_page_index(vma, address);
1046 * __page_check_anon_rmap - sanity check anonymous rmap addition
1047 * @page: the page to add the mapping to
1048 * @vma: the vm area in which the mapping is added
1049 * @address: the user virtual address mapped
1051 static void __page_check_anon_rmap(struct page *page,
1052 struct vm_area_struct *vma, unsigned long address)
1054 #ifdef CONFIG_DEBUG_VM
1056 * The page's anon-rmap details (mapping and index) are guaranteed to
1057 * be set up correctly at this point.
1059 * We have exclusion against page_add_anon_rmap because the caller
1060 * always holds the page locked, except if called from page_dup_rmap,
1061 * in which case the page is already known to be setup.
1063 * We have exclusion against page_add_new_anon_rmap because those pages
1064 * are initially only visible via the pagetables, and the pte is locked
1065 * over the call to page_add_new_anon_rmap.
1067 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1068 BUG_ON(page_to_pgoff(page) != linear_page_index(vma, address));
1073 * page_add_anon_rmap - add pte mapping to an anonymous page
1074 * @page: the page to add the mapping to
1075 * @vma: the vm area in which the mapping is added
1076 * @address: the user virtual address mapped
1077 * @compound: charge the page as compound or small page
1079 * The caller needs to hold the pte lock, and the page must be locked in
1080 * the anon_vma case: to serialize mapping,index checking after setting,
1081 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1082 * (but PageKsm is never downgraded to PageAnon).
1084 void page_add_anon_rmap(struct page *page,
1085 struct vm_area_struct *vma, unsigned long address, bool compound)
1087 do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
1091 * Special version of the above for do_swap_page, which often runs
1092 * into pages that are exclusively owned by the current process.
1093 * Everybody else should continue to use page_add_anon_rmap above.
1095 void do_page_add_anon_rmap(struct page *page,
1096 struct vm_area_struct *vma, unsigned long address, int flags)
1098 bool compound = flags & RMAP_COMPOUND;
1103 VM_BUG_ON_PAGE(!PageLocked(page), page);
1104 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1105 mapcount = compound_mapcount_ptr(page);
1106 first = atomic_inc_and_test(mapcount);
1108 first = atomic_inc_and_test(&page->_mapcount);
1112 int nr = compound ? hpage_nr_pages(page) : 1;
1114 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1115 * these counters are not modified in interrupt context, and
1116 * pte lock(a spinlock) is held, which implies preemption
1120 __inc_node_page_state(page, NR_ANON_THPS);
1121 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1123 if (unlikely(PageKsm(page)))
1126 VM_BUG_ON_PAGE(!PageLocked(page), page);
1128 /* address might be in next vma when migration races vma_adjust */
1130 __page_set_anon_rmap(page, vma, address,
1131 flags & RMAP_EXCLUSIVE);
1133 __page_check_anon_rmap(page, vma, address);
1137 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1138 * @page: the page to add the mapping to
1139 * @vma: the vm area in which the mapping is added
1140 * @address: the user virtual address mapped
1141 * @compound: charge the page as compound or small page
1143 * Same as page_add_anon_rmap but must only be called on *new* pages.
1144 * This means the inc-and-test can be bypassed.
1145 * Page does not have to be locked.
1147 void page_add_new_anon_rmap(struct page *page,
1148 struct vm_area_struct *vma, unsigned long address, bool compound)
1150 int nr = compound ? hpage_nr_pages(page) : 1;
1152 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1153 __SetPageSwapBacked(page);
1155 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1156 /* increment count (starts at -1) */
1157 atomic_set(compound_mapcount_ptr(page), 0);
1158 __inc_node_page_state(page, NR_ANON_THPS);
1160 /* Anon THP always mapped first with PMD */
1161 VM_BUG_ON_PAGE(PageTransCompound(page), page);
1162 /* increment count (starts at -1) */
1163 atomic_set(&page->_mapcount, 0);
1165 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1166 __page_set_anon_rmap(page, vma, address, 1);
1170 * page_add_file_rmap - add pte mapping to a file page
1171 * @page: the page to add the mapping to
1173 * The caller needs to hold the pte lock.
1175 void page_add_file_rmap(struct page *page, bool compound)
1179 VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1180 lock_page_memcg(page);
1181 if (compound && PageTransHuge(page)) {
1182 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1183 if (atomic_inc_and_test(&page[i]._mapcount))
1186 if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1188 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
1189 __inc_node_page_state(page, NR_SHMEM_PMDMAPPED);
1191 if (PageTransCompound(page) && page_mapping(page)) {
1192 VM_WARN_ON_ONCE(!PageLocked(page));
1194 SetPageDoubleMap(compound_head(page));
1195 if (PageMlocked(page))
1196 clear_page_mlock(compound_head(page));
1198 if (!atomic_inc_and_test(&page->_mapcount))
1201 __mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
1203 unlock_page_memcg(page);
1206 static void page_remove_file_rmap(struct page *page, bool compound)
1210 VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1211 lock_page_memcg(page);
1213 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1214 if (unlikely(PageHuge(page))) {
1215 /* hugetlb pages are always mapped with pmds */
1216 atomic_dec(compound_mapcount_ptr(page));
1220 /* page still mapped by someone else? */
1221 if (compound && PageTransHuge(page)) {
1222 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1223 if (atomic_add_negative(-1, &page[i]._mapcount))
1226 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1228 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
1229 __dec_node_page_state(page, NR_SHMEM_PMDMAPPED);
1231 if (!atomic_add_negative(-1, &page->_mapcount))
1236 * We use the irq-unsafe __{inc|mod}_lruvec_page_state because
1237 * these counters are not modified in interrupt context, and
1238 * pte lock(a spinlock) is held, which implies preemption disabled.
1240 __mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr);
1242 if (unlikely(PageMlocked(page)))
1243 clear_page_mlock(page);
1245 unlock_page_memcg(page);
1248 static void page_remove_anon_compound_rmap(struct page *page)
1252 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1255 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1256 if (unlikely(PageHuge(page)))
1259 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1262 __dec_node_page_state(page, NR_ANON_THPS);
1264 if (TestClearPageDoubleMap(page)) {
1266 * Subpages can be mapped with PTEs too. Check how many of
1267 * themi are still mapped.
1269 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1270 if (atomic_add_negative(-1, &page[i]._mapcount))
1277 if (unlikely(PageMlocked(page)))
1278 clear_page_mlock(page);
1281 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, -nr);
1282 deferred_split_huge_page(page);
1287 * page_remove_rmap - take down pte mapping from a page
1288 * @page: page to remove mapping from
1289 * @compound: uncharge the page as compound or small page
1291 * The caller needs to hold the pte lock.
1293 void page_remove_rmap(struct page *page, bool compound)
1295 if (!PageAnon(page))
1296 return page_remove_file_rmap(page, compound);
1299 return page_remove_anon_compound_rmap(page);
1301 /* page still mapped by someone else? */
1302 if (!atomic_add_negative(-1, &page->_mapcount))
1306 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1307 * these counters are not modified in interrupt context, and
1308 * pte lock(a spinlock) is held, which implies preemption disabled.
1310 __dec_node_page_state(page, NR_ANON_MAPPED);
1312 if (unlikely(PageMlocked(page)))
1313 clear_page_mlock(page);
1315 if (PageTransCompound(page))
1316 deferred_split_huge_page(compound_head(page));
1319 * It would be tidy to reset the PageAnon mapping here,
1320 * but that might overwrite a racing page_add_anon_rmap
1321 * which increments mapcount after us but sets mapping
1322 * before us: so leave the reset to free_hot_cold_page,
1323 * and remember that it's only reliable while mapped.
1324 * Leaving it set also helps swapoff to reinstate ptes
1325 * faster for those pages still in swapcache.
1330 * @arg: enum ttu_flags will be passed to this argument
1332 static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1333 unsigned long address, void *arg)
1335 struct mm_struct *mm = vma->vm_mm;
1336 struct page_vma_mapped_walk pvmw = {
1342 struct page *subpage;
1344 unsigned long start = address, end;
1345 enum ttu_flags flags = (enum ttu_flags)arg;
1348 * When racing against e.g. zap_pte_range() on another cpu,
1349 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1350 * try_to_unmap() may return false when it is about to become true,
1351 * if page table locking is skipped: use TTU_SYNC to wait for that.
1353 if (flags & TTU_SYNC)
1354 pvmw.flags = PVMW_SYNC;
1356 /* munlock has nothing to gain from examining un-locked vmas */
1357 if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
1360 if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
1361 is_zone_device_page(page) && !is_device_private_page(page))
1364 if (flags & TTU_SPLIT_HUGE_PMD) {
1365 split_huge_pmd_address(vma, address,
1366 flags & TTU_SPLIT_FREEZE, page);
1370 * For THP, we have to assume the worse case ie pmd for invalidation.
1371 * For hugetlb, it could be much worse if we need to do pud
1372 * invalidation in the case of pmd sharing.
1374 * Note that the page can not be free in this function as call of
1375 * try_to_unmap() must hold a reference on the page.
1377 end = PageKsm(page) ?
1378 address + PAGE_SIZE : vma_address_end(page, vma);
1379 if (PageHuge(page)) {
1381 * If sharing is possible, start and end will be adjusted
1384 adjust_range_if_pmd_sharing_possible(vma, &start, &end);
1386 mmu_notifier_invalidate_range_start(vma->vm_mm, start, end);
1388 while (page_vma_mapped_walk(&pvmw)) {
1389 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1390 /* PMD-mapped THP migration entry */
1391 if (!pvmw.pte && (flags & TTU_MIGRATION)) {
1392 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
1394 if (!PageAnon(page))
1397 set_pmd_migration_entry(&pvmw, page);
1403 * If the page is mlock()d, we cannot swap it out.
1404 * If it's recently referenced (perhaps page_referenced
1405 * skipped over this mm) then we should reactivate it.
1407 if (!(flags & TTU_IGNORE_MLOCK)) {
1408 if (vma->vm_flags & VM_LOCKED) {
1409 /* PTE-mapped THP are never mlocked */
1410 if (!PageTransCompound(page)) {
1412 * Holding pte lock, we do *not* need
1415 mlock_vma_page(page);
1418 page_vma_mapped_walk_done(&pvmw);
1421 if (flags & TTU_MUNLOCK)
1425 /* Unexpected PMD-mapped THP? */
1426 VM_BUG_ON_PAGE(!pvmw.pte, page);
1428 subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
1429 address = pvmw.address;
1431 if (PageHuge(page)) {
1432 if (huge_pmd_unshare(mm, &address, pvmw.pte)) {
1434 * huge_pmd_unshare unmapped an entire PMD
1435 * page. There is no way of knowing exactly
1436 * which PMDs may be cached for this mm, so
1437 * we must flush them all. start/end were
1438 * already adjusted above to cover this range.
1440 flush_cache_range(vma, start, end);
1441 flush_tlb_range(vma, start, end);
1442 mmu_notifier_invalidate_range(mm, start, end);
1445 * The ref count of the PMD page was dropped
1446 * which is part of the way map counting
1447 * is done for shared PMDs. Return 'true'
1448 * here. When there is no other sharing,
1449 * huge_pmd_unshare returns false and we will
1450 * unmap the actual page and drop map count
1453 page_vma_mapped_walk_done(&pvmw);
1458 if (IS_ENABLED(CONFIG_MIGRATION) &&
1459 (flags & TTU_MIGRATION) &&
1460 is_zone_device_page(page)) {
1464 pteval = ptep_get_and_clear(mm, pvmw.address, pvmw.pte);
1467 * Store the pfn of the page in a special migration
1468 * pte. do_swap_page() will wait until the migration
1469 * pte is removed and then restart fault handling.
1471 entry = make_migration_entry(page, 0);
1472 swp_pte = swp_entry_to_pte(entry);
1473 if (pte_soft_dirty(pteval))
1474 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1475 set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
1479 if (!(flags & TTU_IGNORE_ACCESS)) {
1480 if (ptep_clear_flush_young_notify(vma, address,
1483 page_vma_mapped_walk_done(&pvmw);
1488 /* Nuke the page table entry. */
1489 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1490 if (should_defer_flush(mm, flags)) {
1492 * We clear the PTE but do not flush so potentially
1493 * a remote CPU could still be writing to the page.
1494 * If the entry was previously clean then the
1495 * architecture must guarantee that a clear->dirty
1496 * transition on a cached TLB entry is written through
1497 * and traps if the PTE is unmapped.
1499 pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1501 set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1503 pteval = ptep_clear_flush(vma, address, pvmw.pte);
1506 /* Move the dirty bit to the page. Now the pte is gone. */
1507 if (pte_dirty(pteval))
1508 set_page_dirty(page);
1510 /* Update high watermark before we lower rss */
1511 update_hiwater_rss(mm);
1513 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1514 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1515 if (PageHuge(page)) {
1516 int nr = 1 << compound_order(page);
1517 hugetlb_count_sub(nr, mm);
1518 set_huge_swap_pte_at(mm, address,
1520 vma_mmu_pagesize(vma));
1522 dec_mm_counter(mm, mm_counter(page));
1523 set_pte_at(mm, address, pvmw.pte, pteval);
1526 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1528 * The guest indicated that the page content is of no
1529 * interest anymore. Simply discard the pte, vmscan
1530 * will take care of the rest.
1531 * A future reference will then fault in a new zero
1532 * page. When userfaultfd is active, we must not drop
1533 * this page though, as its main user (postcopy
1534 * migration) will not expect userfaults on already
1537 dec_mm_counter(mm, mm_counter(page));
1538 } else if (IS_ENABLED(CONFIG_MIGRATION) &&
1539 (flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))) {
1543 * Store the pfn of the page in a special migration
1544 * pte. do_swap_page() will wait until the migration
1545 * pte is removed and then restart fault handling.
1547 entry = make_migration_entry(subpage,
1549 swp_pte = swp_entry_to_pte(entry);
1550 if (pte_soft_dirty(pteval))
1551 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1552 set_pte_at(mm, address, pvmw.pte, swp_pte);
1553 } else if (PageAnon(page)) {
1554 swp_entry_t entry = { .val = page_private(subpage) };
1557 * Store the swap location in the pte.
1558 * See handle_pte_fault() ...
1560 if (unlikely(PageSwapBacked(page) != PageSwapCache(page))) {
1563 /* We have to invalidate as we cleared the pte */
1564 page_vma_mapped_walk_done(&pvmw);
1568 /* MADV_FREE page check */
1569 if (!PageSwapBacked(page)) {
1570 int ref_count, map_count;
1573 * Synchronize with gup_pte_range():
1574 * - clear PTE; barrier; read refcount
1575 * - inc refcount; barrier; read PTE
1579 ref_count = page_ref_count(page);
1580 map_count = page_mapcount(page);
1583 * Order reads for page refcount and dirty flag
1584 * (see comments in __remove_mapping()).
1589 * The only page refs must be one from isolation
1590 * plus the rmap(s) (dropped by discard:).
1592 if (ref_count == 1 + map_count &&
1594 dec_mm_counter(mm, MM_ANONPAGES);
1599 * If the page was redirtied, it cannot be
1600 * discarded. Remap the page to page table.
1602 set_pte_at(mm, address, pvmw.pte, pteval);
1603 SetPageSwapBacked(page);
1605 page_vma_mapped_walk_done(&pvmw);
1609 if (swap_duplicate(entry) < 0) {
1610 set_pte_at(mm, address, pvmw.pte, pteval);
1612 page_vma_mapped_walk_done(&pvmw);
1615 if (list_empty(&mm->mmlist)) {
1616 spin_lock(&mmlist_lock);
1617 if (list_empty(&mm->mmlist))
1618 list_add(&mm->mmlist, &init_mm.mmlist);
1619 spin_unlock(&mmlist_lock);
1621 dec_mm_counter(mm, MM_ANONPAGES);
1622 inc_mm_counter(mm, MM_SWAPENTS);
1623 swp_pte = swp_entry_to_pte(entry);
1624 if (pte_soft_dirty(pteval))
1625 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1626 set_pte_at(mm, address, pvmw.pte, swp_pte);
1628 dec_mm_counter(mm, mm_counter_file(page));
1630 page_remove_rmap(subpage, PageHuge(page));
1632 mmu_notifier_invalidate_range(mm, address,
1633 address + PAGE_SIZE);
1636 mmu_notifier_invalidate_range_end(vma->vm_mm, start, end);
1641 bool is_vma_temporary_stack(struct vm_area_struct *vma)
1643 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1648 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1649 VM_STACK_INCOMPLETE_SETUP)
1655 static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1657 return is_vma_temporary_stack(vma);
1660 static int page_not_mapped(struct page *page)
1662 return !page_mapped(page);
1666 * try_to_unmap - try to remove all page table mappings to a page
1667 * @page: the page to get unmapped
1668 * @flags: action and flags
1670 * Tries to remove all the page table entries which are mapping this
1671 * page, used in the pageout path. Caller must hold the page lock.
1673 * If unmap is successful, return true. Otherwise, false.
1675 bool try_to_unmap(struct page *page, enum ttu_flags flags)
1677 struct rmap_walk_control rwc = {
1678 .rmap_one = try_to_unmap_one,
1679 .arg = (void *)flags,
1680 .done = page_not_mapped,
1681 .anon_lock = page_lock_anon_vma_read,
1685 * During exec, a temporary VMA is setup and later moved.
1686 * The VMA is moved under the anon_vma lock but not the
1687 * page tables leading to a race where migration cannot
1688 * find the migration ptes. Rather than increasing the
1689 * locking requirements of exec(), migration skips
1690 * temporary VMAs until after exec() completes.
1692 if ((flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))
1693 && !PageKsm(page) && PageAnon(page))
1694 rwc.invalid_vma = invalid_migration_vma;
1696 if (flags & TTU_RMAP_LOCKED)
1697 rmap_walk_locked(page, &rwc);
1699 rmap_walk(page, &rwc);
1702 * When racing against e.g. zap_pte_range() on another cpu,
1703 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1704 * try_to_unmap() may return false when it is about to become true,
1705 * if page table locking is skipped: use TTU_SYNC to wait for that.
1707 return !page_mapcount(page);
1711 * try_to_munlock - try to munlock a page
1712 * @page: the page to be munlocked
1714 * Called from munlock code. Checks all of the VMAs mapping the page
1715 * to make sure nobody else has this page mlocked. The page will be
1716 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1719 void try_to_munlock(struct page *page)
1721 struct rmap_walk_control rwc = {
1722 .rmap_one = try_to_unmap_one,
1723 .arg = (void *)TTU_MUNLOCK,
1724 .done = page_not_mapped,
1725 .anon_lock = page_lock_anon_vma_read,
1729 VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1730 VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
1732 rmap_walk(page, &rwc);
1735 void __put_anon_vma(struct anon_vma *anon_vma)
1737 struct anon_vma *root = anon_vma->root;
1739 anon_vma_free(anon_vma);
1740 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1741 anon_vma_free(root);
1744 static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1745 struct rmap_walk_control *rwc)
1747 struct anon_vma *anon_vma;
1750 return rwc->anon_lock(page);
1753 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1754 * because that depends on page_mapped(); but not all its usages
1755 * are holding mmap_sem. Users without mmap_sem are required to
1756 * take a reference count to prevent the anon_vma disappearing
1758 anon_vma = page_anon_vma(page);
1762 anon_vma_lock_read(anon_vma);
1767 * rmap_walk_anon - do something to anonymous page using the object-based
1769 * @page: the page to be handled
1770 * @rwc: control variable according to each walk type
1772 * Find all the mappings of a page using the mapping pointer and the vma chains
1773 * contained in the anon_vma struct it points to.
1775 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1776 * where the page was found will be held for write. So, we won't recheck
1777 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1780 static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
1783 struct anon_vma *anon_vma;
1784 pgoff_t pgoff_start, pgoff_end;
1785 struct anon_vma_chain *avc;
1788 anon_vma = page_anon_vma(page);
1789 /* anon_vma disappear under us? */
1790 VM_BUG_ON_PAGE(!anon_vma, page);
1792 anon_vma = rmap_walk_anon_lock(page, rwc);
1797 pgoff_start = page_to_pgoff(page);
1798 pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1799 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1800 pgoff_start, pgoff_end) {
1801 struct vm_area_struct *vma = avc->vma;
1802 unsigned long address = vma_address(page, vma);
1804 VM_BUG_ON_VMA(address == -EFAULT, vma);
1807 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1810 if (!rwc->rmap_one(page, vma, address, rwc->arg))
1812 if (rwc->done && rwc->done(page))
1817 anon_vma_unlock_read(anon_vma);
1821 * rmap_walk_file - do something to file page using the object-based rmap method
1822 * @page: the page to be handled
1823 * @rwc: control variable according to each walk type
1825 * Find all the mappings of a page using the mapping pointer and the vma chains
1826 * contained in the address_space struct it points to.
1828 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1829 * where the page was found will be held for write. So, we won't recheck
1830 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1833 static void rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
1836 struct address_space *mapping = page_mapping(page);
1837 pgoff_t pgoff_start, pgoff_end;
1838 struct vm_area_struct *vma;
1841 * The page lock not only makes sure that page->mapping cannot
1842 * suddenly be NULLified by truncation, it makes sure that the
1843 * structure at mapping cannot be freed and reused yet,
1844 * so we can safely take mapping->i_mmap_rwsem.
1846 VM_BUG_ON_PAGE(!PageLocked(page), page);
1851 pgoff_start = page_to_pgoff(page);
1852 pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1854 i_mmap_lock_read(mapping);
1855 vma_interval_tree_foreach(vma, &mapping->i_mmap,
1856 pgoff_start, pgoff_end) {
1857 unsigned long address = vma_address(page, vma);
1859 VM_BUG_ON_VMA(address == -EFAULT, vma);
1862 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1865 if (!rwc->rmap_one(page, vma, address, rwc->arg))
1867 if (rwc->done && rwc->done(page))
1873 i_mmap_unlock_read(mapping);
1876 void rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1878 if (unlikely(PageKsm(page)))
1879 rmap_walk_ksm(page, rwc);
1880 else if (PageAnon(page))
1881 rmap_walk_anon(page, rwc, false);
1883 rmap_walk_file(page, rwc, false);
1886 /* Like rmap_walk, but caller holds relevant rmap lock */
1887 void rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
1889 /* no ksm support for now */
1890 VM_BUG_ON_PAGE(PageKsm(page), page);
1892 rmap_walk_anon(page, rwc, true);
1894 rmap_walk_file(page, rwc, true);
1897 #ifdef CONFIG_HUGETLB_PAGE
1899 * The following three functions are for anonymous (private mapped) hugepages.
1900 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1901 * and no lru code, because we handle hugepages differently from common pages.
1903 static void __hugepage_set_anon_rmap(struct page *page,
1904 struct vm_area_struct *vma, unsigned long address, int exclusive)
1906 struct anon_vma *anon_vma = vma->anon_vma;
1913 anon_vma = anon_vma->root;
1915 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1916 page->mapping = (struct address_space *) anon_vma;
1917 page->index = linear_page_index(vma, address);
1920 void hugepage_add_anon_rmap(struct page *page,
1921 struct vm_area_struct *vma, unsigned long address)
1923 struct anon_vma *anon_vma = vma->anon_vma;
1926 BUG_ON(!PageLocked(page));
1928 /* address might be in next vma when migration races vma_adjust */
1929 first = atomic_inc_and_test(compound_mapcount_ptr(page));
1931 __hugepage_set_anon_rmap(page, vma, address, 0);
1934 void hugepage_add_new_anon_rmap(struct page *page,
1935 struct vm_area_struct *vma, unsigned long address)
1937 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1938 atomic_set(compound_mapcount_ptr(page), 0);
1939 __hugepage_set_anon_rmap(page, vma, address, 1);
1941 #endif /* CONFIG_HUGETLB_PAGE */