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 * i_pages lock (widely used)
36 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
37 * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
38 * sb_lock (within inode_lock in fs/fs-writeback.c)
39 * i_pages lock (widely used, in set_page_dirty,
40 * in arch-dependent flush_dcache_mmap_lock,
41 * within bdi.wb->list_lock in __sync_single_inode)
43 * anon_vma->rwsem,mapping->i_mutex (memory_failure, collect_procs_anon)
49 #include <linux/sched/mm.h>
50 #include <linux/sched/task.h>
51 #include <linux/pagemap.h>
52 #include <linux/swap.h>
53 #include <linux/swapops.h>
54 #include <linux/slab.h>
55 #include <linux/init.h>
56 #include <linux/ksm.h>
57 #include <linux/rmap.h>
58 #include <linux/rcupdate.h>
59 #include <linux/export.h>
60 #include <linux/memcontrol.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/migrate.h>
63 #include <linux/hugetlb.h>
64 #include <linux/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->degree = 1; /* Reference for first vma */
86 anon_vma->parent = anon_vma;
88 * Initialise the anon_vma root to point to itself. If called
89 * from fork, the root will be reset to the parents anon_vma.
91 anon_vma->root = anon_vma;
97 static inline void anon_vma_free(struct anon_vma *anon_vma)
99 VM_BUG_ON(atomic_read(&anon_vma->refcount));
102 * Synchronize against page_lock_anon_vma_read() such that
103 * we can safely hold the lock without the anon_vma getting
106 * Relies on the full mb implied by the atomic_dec_and_test() from
107 * put_anon_vma() against the acquire barrier implied by
108 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
110 * page_lock_anon_vma_read() VS put_anon_vma()
111 * down_read_trylock() atomic_dec_and_test()
113 * atomic_read() rwsem_is_locked()
115 * LOCK should suffice since the actual taking of the lock must
116 * happen _before_ what follows.
119 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
120 anon_vma_lock_write(anon_vma);
121 anon_vma_unlock_write(anon_vma);
124 kmem_cache_free(anon_vma_cachep, anon_vma);
127 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
129 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
132 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
134 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
137 static void anon_vma_chain_link(struct vm_area_struct *vma,
138 struct anon_vma_chain *avc,
139 struct anon_vma *anon_vma)
142 avc->anon_vma = anon_vma;
143 list_add(&avc->same_vma, &vma->anon_vma_chain);
144 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
148 * __anon_vma_prepare - attach an anon_vma to a memory region
149 * @vma: the memory region in question
151 * This makes sure the memory mapping described by 'vma' has
152 * an 'anon_vma' attached to it, so that we can associate the
153 * anonymous pages mapped into it with that anon_vma.
155 * The common case will be that we already have one, which
156 * is handled inline by anon_vma_prepare(). But if
157 * not we either need to find an adjacent mapping that we
158 * can re-use the anon_vma from (very common when the only
159 * reason for splitting a vma has been mprotect()), or we
160 * allocate a new one.
162 * Anon-vma allocations are very subtle, because we may have
163 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
164 * and that may actually touch the spinlock even in the newly
165 * allocated vma (it depends on RCU to make sure that the
166 * anon_vma isn't actually destroyed).
168 * As a result, we need to do proper anon_vma locking even
169 * for the new allocation. At the same time, we do not want
170 * to do any locking for the common case of already having
173 * This must be called with the mmap_sem held for reading.
175 int __anon_vma_prepare(struct vm_area_struct *vma)
177 struct mm_struct *mm = vma->vm_mm;
178 struct anon_vma *anon_vma, *allocated;
179 struct anon_vma_chain *avc;
183 avc = anon_vma_chain_alloc(GFP_KERNEL);
187 anon_vma = find_mergeable_anon_vma(vma);
190 anon_vma = anon_vma_alloc();
191 if (unlikely(!anon_vma))
192 goto out_enomem_free_avc;
193 allocated = anon_vma;
196 anon_vma_lock_write(anon_vma);
197 /* page_table_lock to protect against threads */
198 spin_lock(&mm->page_table_lock);
199 if (likely(!vma->anon_vma)) {
200 vma->anon_vma = anon_vma;
201 anon_vma_chain_link(vma, avc, anon_vma);
202 /* vma reference or self-parent link for new root */
207 spin_unlock(&mm->page_table_lock);
208 anon_vma_unlock_write(anon_vma);
210 if (unlikely(allocated))
211 put_anon_vma(allocated);
213 anon_vma_chain_free(avc);
218 anon_vma_chain_free(avc);
224 * This is a useful helper function for locking the anon_vma root as
225 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
228 * Such anon_vma's should have the same root, so you'd expect to see
229 * just a single mutex_lock for the whole traversal.
231 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
233 struct anon_vma *new_root = anon_vma->root;
234 if (new_root != root) {
235 if (WARN_ON_ONCE(root))
236 up_write(&root->rwsem);
238 down_write(&root->rwsem);
243 static inline void unlock_anon_vma_root(struct anon_vma *root)
246 up_write(&root->rwsem);
250 * Attach the anon_vmas from src to dst.
251 * Returns 0 on success, -ENOMEM on failure.
253 * If dst->anon_vma is NULL this function tries to find and reuse existing
254 * anon_vma which has no vmas and only one child anon_vma. This prevents
255 * degradation of anon_vma hierarchy to endless linear chain in case of
256 * constantly forking task. On the other hand, an anon_vma with more than one
257 * child isn't reused even if there was no alive vma, thus rmap walker has a
258 * good chance of avoiding scanning the whole hierarchy when it searches where
261 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
263 struct anon_vma_chain *avc, *pavc;
264 struct anon_vma *root = NULL;
266 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
267 struct anon_vma *anon_vma;
269 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
270 if (unlikely(!avc)) {
271 unlock_anon_vma_root(root);
273 avc = anon_vma_chain_alloc(GFP_KERNEL);
277 anon_vma = pavc->anon_vma;
278 root = lock_anon_vma_root(root, anon_vma);
279 anon_vma_chain_link(dst, avc, anon_vma);
282 * Reuse existing anon_vma if its degree lower than two,
283 * that means it has no vma and only one anon_vma child.
285 * Do not chose parent anon_vma, otherwise first child
286 * will always reuse it. Root anon_vma is never reused:
287 * it has self-parent reference and at least one child.
289 if (!dst->anon_vma && anon_vma != src->anon_vma &&
290 anon_vma->degree < 2)
291 dst->anon_vma = anon_vma;
294 dst->anon_vma->degree++;
295 unlock_anon_vma_root(root);
300 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
301 * decremented in unlink_anon_vmas().
302 * We can safely do this because callers of anon_vma_clone() don't care
303 * about dst->anon_vma if anon_vma_clone() failed.
305 dst->anon_vma = NULL;
306 unlink_anon_vmas(dst);
311 * Attach vma to its own anon_vma, as well as to the anon_vmas that
312 * the corresponding VMA in the parent process is attached to.
313 * Returns 0 on success, non-zero on failure.
315 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
317 struct anon_vma_chain *avc;
318 struct anon_vma *anon_vma;
321 /* Don't bother if the parent process has no anon_vma here. */
325 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
326 vma->anon_vma = NULL;
329 * First, attach the new VMA to the parent VMA's anon_vmas,
330 * so rmap can find non-COWed pages in child processes.
332 error = anon_vma_clone(vma, pvma);
336 /* An existing anon_vma has been reused, all done then. */
340 /* Then add our own anon_vma. */
341 anon_vma = anon_vma_alloc();
344 avc = anon_vma_chain_alloc(GFP_KERNEL);
346 goto out_error_free_anon_vma;
349 * The root anon_vma's spinlock is the lock actually used when we
350 * lock any of the anon_vmas in this anon_vma tree.
352 anon_vma->root = pvma->anon_vma->root;
353 anon_vma->parent = pvma->anon_vma;
355 * With refcounts, an anon_vma can stay around longer than the
356 * process it belongs to. The root anon_vma needs to be pinned until
357 * this anon_vma is freed, because the lock lives in the root.
359 get_anon_vma(anon_vma->root);
360 /* Mark this anon_vma as the one where our new (COWed) pages go. */
361 vma->anon_vma = anon_vma;
362 anon_vma_lock_write(anon_vma);
363 anon_vma_chain_link(vma, avc, anon_vma);
364 anon_vma->parent->degree++;
365 anon_vma_unlock_write(anon_vma);
369 out_error_free_anon_vma:
370 put_anon_vma(anon_vma);
372 unlink_anon_vmas(vma);
376 void unlink_anon_vmas(struct vm_area_struct *vma)
378 struct anon_vma_chain *avc, *next;
379 struct anon_vma *root = NULL;
382 * Unlink each anon_vma chained to the VMA. This list is ordered
383 * from newest to oldest, ensuring the root anon_vma gets freed last.
385 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
386 struct anon_vma *anon_vma = avc->anon_vma;
388 root = lock_anon_vma_root(root, anon_vma);
389 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
392 * Leave empty anon_vmas on the list - we'll need
393 * to free them outside the lock.
395 if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
396 anon_vma->parent->degree--;
400 list_del(&avc->same_vma);
401 anon_vma_chain_free(avc);
404 vma->anon_vma->degree--;
405 unlock_anon_vma_root(root);
408 * Iterate the list once more, it now only contains empty and unlinked
409 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
410 * needing to write-acquire the anon_vma->root->rwsem.
412 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
413 struct anon_vma *anon_vma = avc->anon_vma;
415 VM_WARN_ON(anon_vma->degree);
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_CACHED;
432 void __init anon_vma_init(void)
434 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
435 0, SLAB_TYPESAFE_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_TYPESAFE_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;
585 if (!tlb_ubc->flush_required)
588 arch_tlbbatch_flush(&tlb_ubc->arch);
589 tlb_ubc->flush_required = false;
590 tlb_ubc->writable = false;
593 /* Flush iff there are potentially writable TLB entries that can race with IO */
594 void try_to_unmap_flush_dirty(void)
596 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
598 if (tlb_ubc->writable)
599 try_to_unmap_flush();
602 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
604 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
606 arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
607 tlb_ubc->flush_required = true;
610 * Ensure compiler does not re-order the setting of tlb_flush_batched
611 * before the PTE is cleared.
614 mm->tlb_flush_batched = true;
617 * If the PTE was dirty then it's best to assume it's writable. The
618 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
619 * before the page is queued for IO.
622 tlb_ubc->writable = true;
626 * Returns true if the TLB flush should be deferred to the end of a batch of
627 * unmap operations to reduce IPIs.
629 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
631 bool should_defer = false;
633 if (!(flags & TTU_BATCH_FLUSH))
636 /* If remote CPUs need to be flushed then defer batch the flush */
637 if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
645 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
646 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
647 * operation such as mprotect or munmap to race between reclaim unmapping
648 * the page and flushing the page. If this race occurs, it potentially allows
649 * access to data via a stale TLB entry. Tracking all mm's that have TLB
650 * batching in flight would be expensive during reclaim so instead track
651 * whether TLB batching occurred in the past and if so then do a flush here
652 * if required. This will cost one additional flush per reclaim cycle paid
653 * by the first operation at risk such as mprotect and mumap.
655 * This must be called under the PTL so that an access to tlb_flush_batched
656 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
659 void flush_tlb_batched_pending(struct mm_struct *mm)
661 if (mm->tlb_flush_batched) {
665 * Do not allow the compiler to re-order the clearing of
666 * tlb_flush_batched before the tlb is flushed.
669 mm->tlb_flush_batched = false;
673 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
677 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
681 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
684 * At what user virtual address is page expected in vma?
685 * Caller should check the page is actually part of the vma.
687 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
689 unsigned long address;
690 if (PageAnon(page)) {
691 struct anon_vma *page__anon_vma = page_anon_vma(page);
693 * Note: swapoff's unuse_vma() is more efficient with this
694 * check, and needs it to match anon_vma when KSM is active.
696 if (!vma->anon_vma || !page__anon_vma ||
697 vma->anon_vma->root != page__anon_vma->root)
699 } else if (page->mapping) {
700 if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping)
704 address = __vma_address(page, vma);
705 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
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 = min(vma->vm_end, start + (PAGE_SIZE << compound_order(page)));
900 mmu_notifier_invalidate_range_start(vma->vm_mm, start, end);
902 while (page_vma_mapped_walk(&pvmw)) {
903 unsigned long cstart;
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);
921 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
922 pmd_t *pmd = pvmw.pmd;
925 if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
928 flush_cache_page(vma, address, page_to_pfn(page));
929 entry = pmdp_huge_clear_flush(vma, address, pmd);
930 entry = pmd_wrprotect(entry);
931 entry = pmd_mkclean(entry);
932 set_pmd_at(vma->vm_mm, address, pmd, entry);
936 /* unexpected pmd-mapped page? */
942 * No need to call mmu_notifier_invalidate_range() as we are
943 * downgrading page table protection not changing it to point
946 * See Documentation/vm/mmu_notifier.rst
952 mmu_notifier_invalidate_range_end(vma->vm_mm, start, end);
957 static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
959 if (vma->vm_flags & VM_SHARED)
965 int page_mkclean(struct page *page)
968 struct address_space *mapping;
969 struct rmap_walk_control rwc = {
970 .arg = (void *)&cleaned,
971 .rmap_one = page_mkclean_one,
972 .invalid_vma = invalid_mkclean_vma,
975 BUG_ON(!PageLocked(page));
977 if (!page_mapped(page))
980 mapping = page_mapping(page);
984 rmap_walk(page, &rwc);
988 EXPORT_SYMBOL_GPL(page_mkclean);
991 * page_move_anon_rmap - move a page to our anon_vma
992 * @page: the page to move to our anon_vma
993 * @vma: the vma the page belongs to
995 * When a page belongs exclusively to one process after a COW event,
996 * that page can be moved into the anon_vma that belongs to just that
997 * process, so the rmap code will not search the parent or sibling
1000 void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
1002 struct anon_vma *anon_vma = vma->anon_vma;
1004 page = compound_head(page);
1006 VM_BUG_ON_PAGE(!PageLocked(page), page);
1007 VM_BUG_ON_VMA(!anon_vma, vma);
1009 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1011 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1012 * simultaneously, so a concurrent reader (eg page_referenced()'s
1013 * PageAnon()) will not see one without the other.
1015 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1019 * __page_set_anon_rmap - set up new anonymous rmap
1020 * @page: Page to add to rmap
1021 * @vma: VM area to add page to.
1022 * @address: User virtual address of the mapping
1023 * @exclusive: the page is exclusively owned by the current process
1025 static void __page_set_anon_rmap(struct page *page,
1026 struct vm_area_struct *vma, unsigned long address, int exclusive)
1028 struct anon_vma *anon_vma = vma->anon_vma;
1036 * If the page isn't exclusively mapped into this vma,
1037 * we must use the _oldest_ possible anon_vma for the
1041 anon_vma = anon_vma->root;
1043 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1044 page->mapping = (struct address_space *) anon_vma;
1045 page->index = linear_page_index(vma, address);
1049 * __page_check_anon_rmap - sanity check anonymous rmap addition
1050 * @page: the page to add the mapping to
1051 * @vma: the vm area in which the mapping is added
1052 * @address: the user virtual address mapped
1054 static void __page_check_anon_rmap(struct page *page,
1055 struct vm_area_struct *vma, unsigned long address)
1057 #ifdef CONFIG_DEBUG_VM
1059 * The page's anon-rmap details (mapping and index) are guaranteed to
1060 * be set up correctly at this point.
1062 * We have exclusion against page_add_anon_rmap because the caller
1063 * always holds the page locked, except if called from page_dup_rmap,
1064 * in which case the page is already known to be setup.
1066 * We have exclusion against page_add_new_anon_rmap because those pages
1067 * are initially only visible via the pagetables, and the pte is locked
1068 * over the call to page_add_new_anon_rmap.
1070 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1071 BUG_ON(page_to_pgoff(page) != linear_page_index(vma, address));
1076 * page_add_anon_rmap - add pte mapping to an anonymous page
1077 * @page: the page to add the mapping to
1078 * @vma: the vm area in which the mapping is added
1079 * @address: the user virtual address mapped
1080 * @compound: charge the page as compound or small page
1082 * The caller needs to hold the pte lock, and the page must be locked in
1083 * the anon_vma case: to serialize mapping,index checking after setting,
1084 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1085 * (but PageKsm is never downgraded to PageAnon).
1087 void page_add_anon_rmap(struct page *page,
1088 struct vm_area_struct *vma, unsigned long address, bool compound)
1090 do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
1094 * Special version of the above for do_swap_page, which often runs
1095 * into pages that are exclusively owned by the current process.
1096 * Everybody else should continue to use page_add_anon_rmap above.
1098 void do_page_add_anon_rmap(struct page *page,
1099 struct vm_area_struct *vma, unsigned long address, int flags)
1101 bool compound = flags & RMAP_COMPOUND;
1106 VM_BUG_ON_PAGE(!PageLocked(page), page);
1107 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1108 mapcount = compound_mapcount_ptr(page);
1109 first = atomic_inc_and_test(mapcount);
1111 first = atomic_inc_and_test(&page->_mapcount);
1115 int nr = compound ? hpage_nr_pages(page) : 1;
1117 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1118 * these counters are not modified in interrupt context, and
1119 * pte lock(a spinlock) is held, which implies preemption
1123 __inc_node_page_state(page, NR_ANON_THPS);
1124 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1126 if (unlikely(PageKsm(page)))
1129 VM_BUG_ON_PAGE(!PageLocked(page), page);
1131 /* address might be in next vma when migration races vma_adjust */
1133 __page_set_anon_rmap(page, vma, address,
1134 flags & RMAP_EXCLUSIVE);
1136 __page_check_anon_rmap(page, vma, address);
1140 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1141 * @page: the page to add the mapping to
1142 * @vma: the vm area in which the mapping is added
1143 * @address: the user virtual address mapped
1144 * @compound: charge the page as compound or small page
1146 * Same as page_add_anon_rmap but must only be called on *new* pages.
1147 * This means the inc-and-test can be bypassed.
1148 * Page does not have to be locked.
1150 void page_add_new_anon_rmap(struct page *page,
1151 struct vm_area_struct *vma, unsigned long address, bool compound)
1153 int nr = compound ? hpage_nr_pages(page) : 1;
1155 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1156 __SetPageSwapBacked(page);
1158 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1159 /* increment count (starts at -1) */
1160 atomic_set(compound_mapcount_ptr(page), 0);
1161 __inc_node_page_state(page, NR_ANON_THPS);
1163 /* Anon THP always mapped first with PMD */
1164 VM_BUG_ON_PAGE(PageTransCompound(page), page);
1165 /* increment count (starts at -1) */
1166 atomic_set(&page->_mapcount, 0);
1168 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1169 __page_set_anon_rmap(page, vma, address, 1);
1173 * page_add_file_rmap - add pte mapping to a file page
1174 * @page: the page to add the mapping to
1175 * @compound: charge the page as compound or small page
1177 * The caller needs to hold the pte lock.
1179 void page_add_file_rmap(struct page *page, bool compound)
1183 VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1184 lock_page_memcg(page);
1185 if (compound && PageTransHuge(page)) {
1186 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1187 if (atomic_inc_and_test(&page[i]._mapcount))
1190 if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1192 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
1193 __inc_node_page_state(page, NR_SHMEM_PMDMAPPED);
1195 if (PageTransCompound(page) && page_mapping(page)) {
1196 VM_WARN_ON_ONCE(!PageLocked(page));
1198 SetPageDoubleMap(compound_head(page));
1199 if (PageMlocked(page))
1200 clear_page_mlock(compound_head(page));
1202 if (!atomic_inc_and_test(&page->_mapcount))
1205 __mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
1207 unlock_page_memcg(page);
1210 static void page_remove_file_rmap(struct page *page, bool compound)
1214 VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1215 lock_page_memcg(page);
1217 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1218 if (unlikely(PageHuge(page))) {
1219 /* hugetlb pages are always mapped with pmds */
1220 atomic_dec(compound_mapcount_ptr(page));
1224 /* page still mapped by someone else? */
1225 if (compound && PageTransHuge(page)) {
1226 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1227 if (atomic_add_negative(-1, &page[i]._mapcount))
1230 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1232 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
1233 __dec_node_page_state(page, NR_SHMEM_PMDMAPPED);
1235 if (!atomic_add_negative(-1, &page->_mapcount))
1240 * We use the irq-unsafe __{inc|mod}_lruvec_page_state because
1241 * these counters are not modified in interrupt context, and
1242 * pte lock(a spinlock) is held, which implies preemption disabled.
1244 __mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr);
1246 if (unlikely(PageMlocked(page)))
1247 clear_page_mlock(page);
1249 unlock_page_memcg(page);
1252 static void page_remove_anon_compound_rmap(struct page *page)
1256 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1259 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1260 if (unlikely(PageHuge(page)))
1263 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1266 __dec_node_page_state(page, NR_ANON_THPS);
1268 if (TestClearPageDoubleMap(page)) {
1270 * Subpages can be mapped with PTEs too. Check how many of
1271 * themi are still mapped.
1273 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1274 if (atomic_add_negative(-1, &page[i]._mapcount))
1281 if (unlikely(PageMlocked(page)))
1282 clear_page_mlock(page);
1285 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, -nr);
1286 deferred_split_huge_page(page);
1291 * page_remove_rmap - take down pte mapping from a page
1292 * @page: page to remove mapping from
1293 * @compound: uncharge the page as compound or small page
1295 * The caller needs to hold the pte lock.
1297 void page_remove_rmap(struct page *page, bool compound)
1299 if (!PageAnon(page))
1300 return page_remove_file_rmap(page, compound);
1303 return page_remove_anon_compound_rmap(page);
1305 /* page still mapped by someone else? */
1306 if (!atomic_add_negative(-1, &page->_mapcount))
1310 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1311 * these counters are not modified in interrupt context, and
1312 * pte lock(a spinlock) is held, which implies preemption disabled.
1314 __dec_node_page_state(page, NR_ANON_MAPPED);
1316 if (unlikely(PageMlocked(page)))
1317 clear_page_mlock(page);
1319 if (PageTransCompound(page))
1320 deferred_split_huge_page(compound_head(page));
1323 * It would be tidy to reset the PageAnon mapping here,
1324 * but that might overwrite a racing page_add_anon_rmap
1325 * which increments mapcount after us but sets mapping
1326 * before us: so leave the reset to free_unref_page,
1327 * and remember that it's only reliable while mapped.
1328 * Leaving it set also helps swapoff to reinstate ptes
1329 * faster for those pages still in swapcache.
1334 * @arg: enum ttu_flags will be passed to this argument
1336 static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1337 unsigned long address, void *arg)
1339 struct mm_struct *mm = vma->vm_mm;
1340 struct page_vma_mapped_walk pvmw = {
1346 struct page *subpage;
1348 unsigned long start = address, end;
1349 enum ttu_flags flags = (enum ttu_flags)arg;
1351 /* munlock has nothing to gain from examining un-locked vmas */
1352 if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
1355 if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
1356 is_zone_device_page(page) && !is_device_private_page(page))
1359 if (flags & TTU_SPLIT_HUGE_PMD) {
1360 split_huge_pmd_address(vma, address,
1361 flags & TTU_SPLIT_FREEZE, page);
1365 * We have to assume the worse case ie pmd for invalidation. Note that
1366 * the page can not be free in this function as call of try_to_unmap()
1367 * must hold a reference on the page.
1369 end = min(vma->vm_end, start + (PAGE_SIZE << compound_order(page)));
1370 mmu_notifier_invalidate_range_start(vma->vm_mm, start, end);
1372 while (page_vma_mapped_walk(&pvmw)) {
1373 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1374 /* PMD-mapped THP migration entry */
1375 if (!pvmw.pte && (flags & TTU_MIGRATION)) {
1376 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
1378 set_pmd_migration_entry(&pvmw, page);
1384 * If the page is mlock()d, we cannot swap it out.
1385 * If it's recently referenced (perhaps page_referenced
1386 * skipped over this mm) then we should reactivate it.
1388 if (!(flags & TTU_IGNORE_MLOCK)) {
1389 if (vma->vm_flags & VM_LOCKED) {
1390 /* PTE-mapped THP are never mlocked */
1391 if (!PageTransCompound(page)) {
1393 * Holding pte lock, we do *not* need
1396 mlock_vma_page(page);
1399 page_vma_mapped_walk_done(&pvmw);
1402 if (flags & TTU_MUNLOCK)
1406 /* Unexpected PMD-mapped THP? */
1407 VM_BUG_ON_PAGE(!pvmw.pte, page);
1409 subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
1410 address = pvmw.address;
1413 if (IS_ENABLED(CONFIG_MIGRATION) &&
1414 (flags & TTU_MIGRATION) &&
1415 is_zone_device_page(page)) {
1419 pteval = ptep_get_and_clear(mm, pvmw.address, pvmw.pte);
1422 * Store the pfn of the page in a special migration
1423 * pte. do_swap_page() will wait until the migration
1424 * pte is removed and then restart fault handling.
1426 entry = make_migration_entry(page, 0);
1427 swp_pte = swp_entry_to_pte(entry);
1428 if (pte_soft_dirty(pteval))
1429 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1430 set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
1432 * No need to invalidate here it will synchronize on
1433 * against the special swap migration pte.
1438 if (!(flags & TTU_IGNORE_ACCESS)) {
1439 if (ptep_clear_flush_young_notify(vma, address,
1442 page_vma_mapped_walk_done(&pvmw);
1447 /* Nuke the page table entry. */
1448 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1449 if (should_defer_flush(mm, flags)) {
1451 * We clear the PTE but do not flush so potentially
1452 * a remote CPU could still be writing to the page.
1453 * If the entry was previously clean then the
1454 * architecture must guarantee that a clear->dirty
1455 * transition on a cached TLB entry is written through
1456 * and traps if the PTE is unmapped.
1458 pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1460 set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1462 pteval = ptep_clear_flush(vma, address, pvmw.pte);
1465 /* Move the dirty bit to the page. Now the pte is gone. */
1466 if (pte_dirty(pteval))
1467 set_page_dirty(page);
1469 /* Update high watermark before we lower rss */
1470 update_hiwater_rss(mm);
1472 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1473 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1474 if (PageHuge(page)) {
1475 int nr = 1 << compound_order(page);
1476 hugetlb_count_sub(nr, mm);
1477 set_huge_swap_pte_at(mm, address,
1479 vma_mmu_pagesize(vma));
1481 dec_mm_counter(mm, mm_counter(page));
1482 set_pte_at(mm, address, pvmw.pte, pteval);
1485 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1487 * The guest indicated that the page content is of no
1488 * interest anymore. Simply discard the pte, vmscan
1489 * will take care of the rest.
1490 * A future reference will then fault in a new zero
1491 * page. When userfaultfd is active, we must not drop
1492 * this page though, as its main user (postcopy
1493 * migration) will not expect userfaults on already
1496 dec_mm_counter(mm, mm_counter(page));
1497 /* We have to invalidate as we cleared the pte */
1498 mmu_notifier_invalidate_range(mm, address,
1499 address + PAGE_SIZE);
1500 } else if (IS_ENABLED(CONFIG_MIGRATION) &&
1501 (flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))) {
1505 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1506 set_pte_at(mm, address, pvmw.pte, pteval);
1508 page_vma_mapped_walk_done(&pvmw);
1513 * Store the pfn of the page in a special migration
1514 * pte. do_swap_page() will wait until the migration
1515 * pte is removed and then restart fault handling.
1517 entry = make_migration_entry(subpage,
1519 swp_pte = swp_entry_to_pte(entry);
1520 if (pte_soft_dirty(pteval))
1521 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1522 set_pte_at(mm, address, pvmw.pte, swp_pte);
1524 * No need to invalidate here it will synchronize on
1525 * against the special swap migration pte.
1527 } else if (PageAnon(page)) {
1528 swp_entry_t entry = { .val = page_private(subpage) };
1531 * Store the swap location in the pte.
1532 * See handle_pte_fault() ...
1534 if (unlikely(PageSwapBacked(page) != PageSwapCache(page))) {
1537 /* We have to invalidate as we cleared the pte */
1538 mmu_notifier_invalidate_range(mm, address,
1539 address + PAGE_SIZE);
1540 page_vma_mapped_walk_done(&pvmw);
1544 /* MADV_FREE page check */
1545 if (!PageSwapBacked(page)) {
1546 if (!PageDirty(page)) {
1547 /* Invalidate as we cleared the pte */
1548 mmu_notifier_invalidate_range(mm,
1549 address, address + PAGE_SIZE);
1550 dec_mm_counter(mm, MM_ANONPAGES);
1555 * If the page was redirtied, it cannot be
1556 * discarded. Remap the page to page table.
1558 set_pte_at(mm, address, pvmw.pte, pteval);
1559 SetPageSwapBacked(page);
1561 page_vma_mapped_walk_done(&pvmw);
1565 if (swap_duplicate(entry) < 0) {
1566 set_pte_at(mm, address, pvmw.pte, pteval);
1568 page_vma_mapped_walk_done(&pvmw);
1571 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1572 set_pte_at(mm, address, pvmw.pte, pteval);
1574 page_vma_mapped_walk_done(&pvmw);
1577 if (list_empty(&mm->mmlist)) {
1578 spin_lock(&mmlist_lock);
1579 if (list_empty(&mm->mmlist))
1580 list_add(&mm->mmlist, &init_mm.mmlist);
1581 spin_unlock(&mmlist_lock);
1583 dec_mm_counter(mm, MM_ANONPAGES);
1584 inc_mm_counter(mm, MM_SWAPENTS);
1585 swp_pte = swp_entry_to_pte(entry);
1586 if (pte_soft_dirty(pteval))
1587 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1588 set_pte_at(mm, address, pvmw.pte, swp_pte);
1589 /* Invalidate as we cleared the pte */
1590 mmu_notifier_invalidate_range(mm, address,
1591 address + PAGE_SIZE);
1594 * We should not need to notify here as we reach this
1595 * case only from freeze_page() itself only call from
1596 * split_huge_page_to_list() so everything below must
1598 * - page is not anonymous
1601 * So as it is a locked file back page thus it can not
1602 * be remove from the page cache and replace by a new
1603 * page before mmu_notifier_invalidate_range_end so no
1604 * concurrent thread might update its page table to
1605 * point at new page while a device still is using this
1608 * See Documentation/vm/mmu_notifier.rst
1610 dec_mm_counter(mm, mm_counter_file(page));
1614 * No need to call mmu_notifier_invalidate_range() it has be
1615 * done above for all cases requiring it to happen under page
1616 * table lock before mmu_notifier_invalidate_range_end()
1618 * See Documentation/vm/mmu_notifier.rst
1620 page_remove_rmap(subpage, PageHuge(page));
1624 mmu_notifier_invalidate_range_end(vma->vm_mm, start, end);
1629 bool is_vma_temporary_stack(struct vm_area_struct *vma)
1631 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1636 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1637 VM_STACK_INCOMPLETE_SETUP)
1643 static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1645 return is_vma_temporary_stack(vma);
1648 static int page_mapcount_is_zero(struct page *page)
1650 return !total_mapcount(page);
1654 * try_to_unmap - try to remove all page table mappings to a page
1655 * @page: the page to get unmapped
1656 * @flags: action and flags
1658 * Tries to remove all the page table entries which are mapping this
1659 * page, used in the pageout path. Caller must hold the page lock.
1661 * If unmap is successful, return true. Otherwise, false.
1663 bool try_to_unmap(struct page *page, enum ttu_flags flags)
1665 struct rmap_walk_control rwc = {
1666 .rmap_one = try_to_unmap_one,
1667 .arg = (void *)flags,
1668 .done = page_mapcount_is_zero,
1669 .anon_lock = page_lock_anon_vma_read,
1673 * During exec, a temporary VMA is setup and later moved.
1674 * The VMA is moved under the anon_vma lock but not the
1675 * page tables leading to a race where migration cannot
1676 * find the migration ptes. Rather than increasing the
1677 * locking requirements of exec(), migration skips
1678 * temporary VMAs until after exec() completes.
1680 if ((flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))
1681 && !PageKsm(page) && PageAnon(page))
1682 rwc.invalid_vma = invalid_migration_vma;
1684 if (flags & TTU_RMAP_LOCKED)
1685 rmap_walk_locked(page, &rwc);
1687 rmap_walk(page, &rwc);
1689 return !page_mapcount(page) ? true : false;
1692 static int page_not_mapped(struct page *page)
1694 return !page_mapped(page);
1698 * try_to_munlock - try to munlock a page
1699 * @page: the page to be munlocked
1701 * Called from munlock code. Checks all of the VMAs mapping the page
1702 * to make sure nobody else has this page mlocked. The page will be
1703 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1706 void try_to_munlock(struct page *page)
1708 struct rmap_walk_control rwc = {
1709 .rmap_one = try_to_unmap_one,
1710 .arg = (void *)TTU_MUNLOCK,
1711 .done = page_not_mapped,
1712 .anon_lock = page_lock_anon_vma_read,
1716 VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1717 VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
1719 rmap_walk(page, &rwc);
1722 void __put_anon_vma(struct anon_vma *anon_vma)
1724 struct anon_vma *root = anon_vma->root;
1726 anon_vma_free(anon_vma);
1727 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1728 anon_vma_free(root);
1731 static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1732 struct rmap_walk_control *rwc)
1734 struct anon_vma *anon_vma;
1737 return rwc->anon_lock(page);
1740 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1741 * because that depends on page_mapped(); but not all its usages
1742 * are holding mmap_sem. Users without mmap_sem are required to
1743 * take a reference count to prevent the anon_vma disappearing
1745 anon_vma = page_anon_vma(page);
1749 anon_vma_lock_read(anon_vma);
1754 * rmap_walk_anon - do something to anonymous page using the object-based
1756 * @page: the page to be handled
1757 * @rwc: control variable according to each walk type
1759 * Find all the mappings of a page using the mapping pointer and the vma chains
1760 * contained in the anon_vma struct it points to.
1762 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1763 * where the page was found will be held for write. So, we won't recheck
1764 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1767 static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
1770 struct anon_vma *anon_vma;
1771 pgoff_t pgoff_start, pgoff_end;
1772 struct anon_vma_chain *avc;
1775 anon_vma = page_anon_vma(page);
1776 /* anon_vma disappear under us? */
1777 VM_BUG_ON_PAGE(!anon_vma, page);
1779 anon_vma = rmap_walk_anon_lock(page, rwc);
1784 pgoff_start = page_to_pgoff(page);
1785 pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1786 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1787 pgoff_start, pgoff_end) {
1788 struct vm_area_struct *vma = avc->vma;
1789 unsigned long address = vma_address(page, vma);
1793 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1796 if (!rwc->rmap_one(page, vma, address, rwc->arg))
1798 if (rwc->done && rwc->done(page))
1803 anon_vma_unlock_read(anon_vma);
1807 * rmap_walk_file - do something to file page using the object-based rmap method
1808 * @page: the page to be handled
1809 * @rwc: control variable according to each walk type
1811 * Find all the mappings of a page using the mapping pointer and the vma chains
1812 * contained in the address_space struct it points to.
1814 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1815 * where the page was found will be held for write. So, we won't recheck
1816 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1819 static void rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
1822 struct address_space *mapping = page_mapping(page);
1823 pgoff_t pgoff_start, pgoff_end;
1824 struct vm_area_struct *vma;
1827 * The page lock not only makes sure that page->mapping cannot
1828 * suddenly be NULLified by truncation, it makes sure that the
1829 * structure at mapping cannot be freed and reused yet,
1830 * so we can safely take mapping->i_mmap_rwsem.
1832 VM_BUG_ON_PAGE(!PageLocked(page), page);
1837 pgoff_start = page_to_pgoff(page);
1838 pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1840 i_mmap_lock_read(mapping);
1841 vma_interval_tree_foreach(vma, &mapping->i_mmap,
1842 pgoff_start, pgoff_end) {
1843 unsigned long address = vma_address(page, vma);
1847 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1850 if (!rwc->rmap_one(page, vma, address, rwc->arg))
1852 if (rwc->done && rwc->done(page))
1858 i_mmap_unlock_read(mapping);
1861 void rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1863 if (unlikely(PageKsm(page)))
1864 rmap_walk_ksm(page, rwc);
1865 else if (PageAnon(page))
1866 rmap_walk_anon(page, rwc, false);
1868 rmap_walk_file(page, rwc, false);
1871 /* Like rmap_walk, but caller holds relevant rmap lock */
1872 void rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
1874 /* no ksm support for now */
1875 VM_BUG_ON_PAGE(PageKsm(page), page);
1877 rmap_walk_anon(page, rwc, true);
1879 rmap_walk_file(page, rwc, true);
1882 #ifdef CONFIG_HUGETLB_PAGE
1884 * The following three functions are for anonymous (private mapped) hugepages.
1885 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1886 * and no lru code, because we handle hugepages differently from common pages.
1888 static void __hugepage_set_anon_rmap(struct page *page,
1889 struct vm_area_struct *vma, unsigned long address, int exclusive)
1891 struct anon_vma *anon_vma = vma->anon_vma;
1898 anon_vma = anon_vma->root;
1900 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1901 page->mapping = (struct address_space *) anon_vma;
1902 page->index = linear_page_index(vma, address);
1905 void hugepage_add_anon_rmap(struct page *page,
1906 struct vm_area_struct *vma, unsigned long address)
1908 struct anon_vma *anon_vma = vma->anon_vma;
1911 BUG_ON(!PageLocked(page));
1913 /* address might be in next vma when migration races vma_adjust */
1914 first = atomic_inc_and_test(compound_mapcount_ptr(page));
1916 __hugepage_set_anon_rmap(page, vma, address, 0);
1919 void hugepage_add_new_anon_rmap(struct page *page,
1920 struct vm_area_struct *vma, unsigned long address)
1922 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1923 atomic_set(compound_mapcount_ptr(page), 0);
1924 __hugepage_set_anon_rmap(page, vma, address, 1);
1926 #endif /* CONFIG_HUGETLB_PAGE */