Back to home page

LXR

 
 

    


0001 /*
0002  * mm/rmap.c - physical to virtual reverse mappings
0003  *
0004  * Copyright 2001, Rik van Riel <riel@conectiva.com.br>
0005  * Released under the General Public License (GPL).
0006  *
0007  * Simple, low overhead reverse mapping scheme.
0008  * Please try to keep this thing as modular as possible.
0009  *
0010  * Provides methods for unmapping each kind of mapped page:
0011  * the anon methods track anonymous pages, and
0012  * the file methods track pages belonging to an inode.
0013  *
0014  * Original design by Rik van Riel <riel@conectiva.com.br> 2001
0015  * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004
0016  * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004
0017  * Contributions by Hugh Dickins 2003, 2004
0018  */
0019 
0020 /*
0021  * Lock ordering in mm:
0022  *
0023  * inode->i_mutex   (while writing or truncating, not reading or faulting)
0024  *   mm->mmap_sem
0025  *     page->flags PG_locked (lock_page)
0026  *       hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share)
0027  *         mapping->i_mmap_rwsem
0028  *           anon_vma->rwsem
0029  *             mm->page_table_lock or pte_lock
0030  *               zone_lru_lock (in mark_page_accessed, isolate_lru_page)
0031  *               swap_lock (in swap_duplicate, swap_info_get)
0032  *                 mmlist_lock (in mmput, drain_mmlist and others)
0033  *                 mapping->private_lock (in __set_page_dirty_buffers)
0034  *                   mem_cgroup_{begin,end}_page_stat (memcg->move_lock)
0035  *                     mapping->tree_lock (widely used)
0036  *                 inode->i_lock (in set_page_dirty's __mark_inode_dirty)
0037  *                 bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
0038  *                   sb_lock (within inode_lock in fs/fs-writeback.c)
0039  *                   mapping->tree_lock (widely used, in set_page_dirty,
0040  *                             in arch-dependent flush_dcache_mmap_lock,
0041  *                             within bdi.wb->list_lock in __sync_single_inode)
0042  *
0043  * anon_vma->rwsem,mapping->i_mutex      (memory_failure, collect_procs_anon)
0044  *   ->tasklist_lock
0045  *     pte map lock
0046  */
0047 
0048 #include <linux/mm.h>
0049 #include <linux/pagemap.h>
0050 #include <linux/swap.h>
0051 #include <linux/swapops.h>
0052 #include <linux/slab.h>
0053 #include <linux/init.h>
0054 #include <linux/ksm.h>
0055 #include <linux/rmap.h>
0056 #include <linux/rcupdate.h>
0057 #include <linux/export.h>
0058 #include <linux/memcontrol.h>
0059 #include <linux/mmu_notifier.h>
0060 #include <linux/migrate.h>
0061 #include <linux/hugetlb.h>
0062 #include <linux/backing-dev.h>
0063 #include <linux/page_idle.h>
0064 
0065 #include <asm/tlbflush.h>
0066 
0067 #include <trace/events/tlb.h>
0068 
0069 #include "internal.h"
0070 
0071 static struct kmem_cache *anon_vma_cachep;
0072 static struct kmem_cache *anon_vma_chain_cachep;
0073 
0074 static inline struct anon_vma *anon_vma_alloc(void)
0075 {
0076     struct anon_vma *anon_vma;
0077 
0078     anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
0079     if (anon_vma) {
0080         atomic_set(&anon_vma->refcount, 1);
0081         anon_vma->degree = 1;   /* Reference for first vma */
0082         anon_vma->parent = anon_vma;
0083         /*
0084          * Initialise the anon_vma root to point to itself. If called
0085          * from fork, the root will be reset to the parents anon_vma.
0086          */
0087         anon_vma->root = anon_vma;
0088     }
0089 
0090     return anon_vma;
0091 }
0092 
0093 static inline void anon_vma_free(struct anon_vma *anon_vma)
0094 {
0095     VM_BUG_ON(atomic_read(&anon_vma->refcount));
0096 
0097     /*
0098      * Synchronize against page_lock_anon_vma_read() such that
0099      * we can safely hold the lock without the anon_vma getting
0100      * freed.
0101      *
0102      * Relies on the full mb implied by the atomic_dec_and_test() from
0103      * put_anon_vma() against the acquire barrier implied by
0104      * down_read_trylock() from page_lock_anon_vma_read(). This orders:
0105      *
0106      * page_lock_anon_vma_read()    VS  put_anon_vma()
0107      *   down_read_trylock()          atomic_dec_and_test()
0108      *   LOCK                 MB
0109      *   atomic_read()            rwsem_is_locked()
0110      *
0111      * LOCK should suffice since the actual taking of the lock must
0112      * happen _before_ what follows.
0113      */
0114     might_sleep();
0115     if (rwsem_is_locked(&anon_vma->root->rwsem)) {
0116         anon_vma_lock_write(anon_vma);
0117         anon_vma_unlock_write(anon_vma);
0118     }
0119 
0120     kmem_cache_free(anon_vma_cachep, anon_vma);
0121 }
0122 
0123 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
0124 {
0125     return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
0126 }
0127 
0128 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
0129 {
0130     kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
0131 }
0132 
0133 static void anon_vma_chain_link(struct vm_area_struct *vma,
0134                 struct anon_vma_chain *avc,
0135                 struct anon_vma *anon_vma)
0136 {
0137     avc->vma = vma;
0138     avc->anon_vma = anon_vma;
0139     list_add(&avc->same_vma, &vma->anon_vma_chain);
0140     anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
0141 }
0142 
0143 /**
0144  * __anon_vma_prepare - attach an anon_vma to a memory region
0145  * @vma: the memory region in question
0146  *
0147  * This makes sure the memory mapping described by 'vma' has
0148  * an 'anon_vma' attached to it, so that we can associate the
0149  * anonymous pages mapped into it with that anon_vma.
0150  *
0151  * The common case will be that we already have one, which
0152  * is handled inline by anon_vma_prepare(). But if
0153  * not we either need to find an adjacent mapping that we
0154  * can re-use the anon_vma from (very common when the only
0155  * reason for splitting a vma has been mprotect()), or we
0156  * allocate a new one.
0157  *
0158  * Anon-vma allocations are very subtle, because we may have
0159  * optimistically looked up an anon_vma in page_lock_anon_vma_read()
0160  * and that may actually touch the spinlock even in the newly
0161  * allocated vma (it depends on RCU to make sure that the
0162  * anon_vma isn't actually destroyed).
0163  *
0164  * As a result, we need to do proper anon_vma locking even
0165  * for the new allocation. At the same time, we do not want
0166  * to do any locking for the common case of already having
0167  * an anon_vma.
0168  *
0169  * This must be called with the mmap_sem held for reading.
0170  */
0171 int __anon_vma_prepare(struct vm_area_struct *vma)
0172 {
0173     struct mm_struct *mm = vma->vm_mm;
0174     struct anon_vma *anon_vma, *allocated;
0175     struct anon_vma_chain *avc;
0176 
0177     might_sleep();
0178 
0179     avc = anon_vma_chain_alloc(GFP_KERNEL);
0180     if (!avc)
0181         goto out_enomem;
0182 
0183     anon_vma = find_mergeable_anon_vma(vma);
0184     allocated = NULL;
0185     if (!anon_vma) {
0186         anon_vma = anon_vma_alloc();
0187         if (unlikely(!anon_vma))
0188             goto out_enomem_free_avc;
0189         allocated = anon_vma;
0190     }
0191 
0192     anon_vma_lock_write(anon_vma);
0193     /* page_table_lock to protect against threads */
0194     spin_lock(&mm->page_table_lock);
0195     if (likely(!vma->anon_vma)) {
0196         vma->anon_vma = anon_vma;
0197         anon_vma_chain_link(vma, avc, anon_vma);
0198         /* vma reference or self-parent link for new root */
0199         anon_vma->degree++;
0200         allocated = NULL;
0201         avc = NULL;
0202     }
0203     spin_unlock(&mm->page_table_lock);
0204     anon_vma_unlock_write(anon_vma);
0205 
0206     if (unlikely(allocated))
0207         put_anon_vma(allocated);
0208     if (unlikely(avc))
0209         anon_vma_chain_free(avc);
0210 
0211     return 0;
0212 
0213  out_enomem_free_avc:
0214     anon_vma_chain_free(avc);
0215  out_enomem:
0216     return -ENOMEM;
0217 }
0218 
0219 /*
0220  * This is a useful helper function for locking the anon_vma root as
0221  * we traverse the vma->anon_vma_chain, looping over anon_vma's that
0222  * have the same vma.
0223  *
0224  * Such anon_vma's should have the same root, so you'd expect to see
0225  * just a single mutex_lock for the whole traversal.
0226  */
0227 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
0228 {
0229     struct anon_vma *new_root = anon_vma->root;
0230     if (new_root != root) {
0231         if (WARN_ON_ONCE(root))
0232             up_write(&root->rwsem);
0233         root = new_root;
0234         down_write(&root->rwsem);
0235     }
0236     return root;
0237 }
0238 
0239 static inline void unlock_anon_vma_root(struct anon_vma *root)
0240 {
0241     if (root)
0242         up_write(&root->rwsem);
0243 }
0244 
0245 /*
0246  * Attach the anon_vmas from src to dst.
0247  * Returns 0 on success, -ENOMEM on failure.
0248  *
0249  * If dst->anon_vma is NULL this function tries to find and reuse existing
0250  * anon_vma which has no vmas and only one child anon_vma. This prevents
0251  * degradation of anon_vma hierarchy to endless linear chain in case of
0252  * constantly forking task. On the other hand, an anon_vma with more than one
0253  * child isn't reused even if there was no alive vma, thus rmap walker has a
0254  * good chance of avoiding scanning the whole hierarchy when it searches where
0255  * page is mapped.
0256  */
0257 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
0258 {
0259     struct anon_vma_chain *avc, *pavc;
0260     struct anon_vma *root = NULL;
0261 
0262     list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
0263         struct anon_vma *anon_vma;
0264 
0265         avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
0266         if (unlikely(!avc)) {
0267             unlock_anon_vma_root(root);
0268             root = NULL;
0269             avc = anon_vma_chain_alloc(GFP_KERNEL);
0270             if (!avc)
0271                 goto enomem_failure;
0272         }
0273         anon_vma = pavc->anon_vma;
0274         root = lock_anon_vma_root(root, anon_vma);
0275         anon_vma_chain_link(dst, avc, anon_vma);
0276 
0277         /*
0278          * Reuse existing anon_vma if its degree lower than two,
0279          * that means it has no vma and only one anon_vma child.
0280          *
0281          * Do not chose parent anon_vma, otherwise first child
0282          * will always reuse it. Root anon_vma is never reused:
0283          * it has self-parent reference and at least one child.
0284          */
0285         if (!dst->anon_vma && anon_vma != src->anon_vma &&
0286                 anon_vma->degree < 2)
0287             dst->anon_vma = anon_vma;
0288     }
0289     if (dst->anon_vma)
0290         dst->anon_vma->degree++;
0291     unlock_anon_vma_root(root);
0292     return 0;
0293 
0294  enomem_failure:
0295     /*
0296      * dst->anon_vma is dropped here otherwise its degree can be incorrectly
0297      * decremented in unlink_anon_vmas().
0298      * We can safely do this because callers of anon_vma_clone() don't care
0299      * about dst->anon_vma if anon_vma_clone() failed.
0300      */
0301     dst->anon_vma = NULL;
0302     unlink_anon_vmas(dst);
0303     return -ENOMEM;
0304 }
0305 
0306 /*
0307  * Attach vma to its own anon_vma, as well as to the anon_vmas that
0308  * the corresponding VMA in the parent process is attached to.
0309  * Returns 0 on success, non-zero on failure.
0310  */
0311 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
0312 {
0313     struct anon_vma_chain *avc;
0314     struct anon_vma *anon_vma;
0315     int error;
0316 
0317     /* Don't bother if the parent process has no anon_vma here. */
0318     if (!pvma->anon_vma)
0319         return 0;
0320 
0321     /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
0322     vma->anon_vma = NULL;
0323 
0324     /*
0325      * First, attach the new VMA to the parent VMA's anon_vmas,
0326      * so rmap can find non-COWed pages in child processes.
0327      */
0328     error = anon_vma_clone(vma, pvma);
0329     if (error)
0330         return error;
0331 
0332     /* An existing anon_vma has been reused, all done then. */
0333     if (vma->anon_vma)
0334         return 0;
0335 
0336     /* Then add our own anon_vma. */
0337     anon_vma = anon_vma_alloc();
0338     if (!anon_vma)
0339         goto out_error;
0340     avc = anon_vma_chain_alloc(GFP_KERNEL);
0341     if (!avc)
0342         goto out_error_free_anon_vma;
0343 
0344     /*
0345      * The root anon_vma's spinlock is the lock actually used when we
0346      * lock any of the anon_vmas in this anon_vma tree.
0347      */
0348     anon_vma->root = pvma->anon_vma->root;
0349     anon_vma->parent = pvma->anon_vma;
0350     /*
0351      * With refcounts, an anon_vma can stay around longer than the
0352      * process it belongs to. The root anon_vma needs to be pinned until
0353      * this anon_vma is freed, because the lock lives in the root.
0354      */
0355     get_anon_vma(anon_vma->root);
0356     /* Mark this anon_vma as the one where our new (COWed) pages go. */
0357     vma->anon_vma = anon_vma;
0358     anon_vma_lock_write(anon_vma);
0359     anon_vma_chain_link(vma, avc, anon_vma);
0360     anon_vma->parent->degree++;
0361     anon_vma_unlock_write(anon_vma);
0362 
0363     return 0;
0364 
0365  out_error_free_anon_vma:
0366     put_anon_vma(anon_vma);
0367  out_error:
0368     unlink_anon_vmas(vma);
0369     return -ENOMEM;
0370 }
0371 
0372 void unlink_anon_vmas(struct vm_area_struct *vma)
0373 {
0374     struct anon_vma_chain *avc, *next;
0375     struct anon_vma *root = NULL;
0376 
0377     /*
0378      * Unlink each anon_vma chained to the VMA.  This list is ordered
0379      * from newest to oldest, ensuring the root anon_vma gets freed last.
0380      */
0381     list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
0382         struct anon_vma *anon_vma = avc->anon_vma;
0383 
0384         root = lock_anon_vma_root(root, anon_vma);
0385         anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
0386 
0387         /*
0388          * Leave empty anon_vmas on the list - we'll need
0389          * to free them outside the lock.
0390          */
0391         if (RB_EMPTY_ROOT(&anon_vma->rb_root)) {
0392             anon_vma->parent->degree--;
0393             continue;
0394         }
0395 
0396         list_del(&avc->same_vma);
0397         anon_vma_chain_free(avc);
0398     }
0399     if (vma->anon_vma)
0400         vma->anon_vma->degree--;
0401     unlock_anon_vma_root(root);
0402 
0403     /*
0404      * Iterate the list once more, it now only contains empty and unlinked
0405      * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
0406      * needing to write-acquire the anon_vma->root->rwsem.
0407      */
0408     list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
0409         struct anon_vma *anon_vma = avc->anon_vma;
0410 
0411         VM_WARN_ON(anon_vma->degree);
0412         put_anon_vma(anon_vma);
0413 
0414         list_del(&avc->same_vma);
0415         anon_vma_chain_free(avc);
0416     }
0417 }
0418 
0419 static void anon_vma_ctor(void *data)
0420 {
0421     struct anon_vma *anon_vma = data;
0422 
0423     init_rwsem(&anon_vma->rwsem);
0424     atomic_set(&anon_vma->refcount, 0);
0425     anon_vma->rb_root = RB_ROOT;
0426 }
0427 
0428 void __init anon_vma_init(void)
0429 {
0430     anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
0431             0, SLAB_DESTROY_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
0432             anon_vma_ctor);
0433     anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
0434             SLAB_PANIC|SLAB_ACCOUNT);
0435 }
0436 
0437 /*
0438  * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
0439  *
0440  * Since there is no serialization what so ever against page_remove_rmap()
0441  * the best this function can do is return a locked anon_vma that might
0442  * have been relevant to this page.
0443  *
0444  * The page might have been remapped to a different anon_vma or the anon_vma
0445  * returned may already be freed (and even reused).
0446  *
0447  * In case it was remapped to a different anon_vma, the new anon_vma will be a
0448  * child of the old anon_vma, and the anon_vma lifetime rules will therefore
0449  * ensure that any anon_vma obtained from the page will still be valid for as
0450  * long as we observe page_mapped() [ hence all those page_mapped() tests ].
0451  *
0452  * All users of this function must be very careful when walking the anon_vma
0453  * chain and verify that the page in question is indeed mapped in it
0454  * [ something equivalent to page_mapped_in_vma() ].
0455  *
0456  * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
0457  * that the anon_vma pointer from page->mapping is valid if there is a
0458  * mapcount, we can dereference the anon_vma after observing those.
0459  */
0460 struct anon_vma *page_get_anon_vma(struct page *page)
0461 {
0462     struct anon_vma *anon_vma = NULL;
0463     unsigned long anon_mapping;
0464 
0465     rcu_read_lock();
0466     anon_mapping = (unsigned long)READ_ONCE(page->mapping);
0467     if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
0468         goto out;
0469     if (!page_mapped(page))
0470         goto out;
0471 
0472     anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
0473     if (!atomic_inc_not_zero(&anon_vma->refcount)) {
0474         anon_vma = NULL;
0475         goto out;
0476     }
0477 
0478     /*
0479      * If this page is still mapped, then its anon_vma cannot have been
0480      * freed.  But if it has been unmapped, we have no security against the
0481      * anon_vma structure being freed and reused (for another anon_vma:
0482      * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
0483      * above cannot corrupt).
0484      */
0485     if (!page_mapped(page)) {
0486         rcu_read_unlock();
0487         put_anon_vma(anon_vma);
0488         return NULL;
0489     }
0490 out:
0491     rcu_read_unlock();
0492 
0493     return anon_vma;
0494 }
0495 
0496 /*
0497  * Similar to page_get_anon_vma() except it locks the anon_vma.
0498  *
0499  * Its a little more complex as it tries to keep the fast path to a single
0500  * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
0501  * reference like with page_get_anon_vma() and then block on the mutex.
0502  */
0503 struct anon_vma *page_lock_anon_vma_read(struct page *page)
0504 {
0505     struct anon_vma *anon_vma = NULL;
0506     struct anon_vma *root_anon_vma;
0507     unsigned long anon_mapping;
0508 
0509     rcu_read_lock();
0510     anon_mapping = (unsigned long)READ_ONCE(page->mapping);
0511     if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
0512         goto out;
0513     if (!page_mapped(page))
0514         goto out;
0515 
0516     anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
0517     root_anon_vma = READ_ONCE(anon_vma->root);
0518     if (down_read_trylock(&root_anon_vma->rwsem)) {
0519         /*
0520          * If the page is still mapped, then this anon_vma is still
0521          * its anon_vma, and holding the mutex ensures that it will
0522          * not go away, see anon_vma_free().
0523          */
0524         if (!page_mapped(page)) {
0525             up_read(&root_anon_vma->rwsem);
0526             anon_vma = NULL;
0527         }
0528         goto out;
0529     }
0530 
0531     /* trylock failed, we got to sleep */
0532     if (!atomic_inc_not_zero(&anon_vma->refcount)) {
0533         anon_vma = NULL;
0534         goto out;
0535     }
0536 
0537     if (!page_mapped(page)) {
0538         rcu_read_unlock();
0539         put_anon_vma(anon_vma);
0540         return NULL;
0541     }
0542 
0543     /* we pinned the anon_vma, its safe to sleep */
0544     rcu_read_unlock();
0545     anon_vma_lock_read(anon_vma);
0546 
0547     if (atomic_dec_and_test(&anon_vma->refcount)) {
0548         /*
0549          * Oops, we held the last refcount, release the lock
0550          * and bail -- can't simply use put_anon_vma() because
0551          * we'll deadlock on the anon_vma_lock_write() recursion.
0552          */
0553         anon_vma_unlock_read(anon_vma);
0554         __put_anon_vma(anon_vma);
0555         anon_vma = NULL;
0556     }
0557 
0558     return anon_vma;
0559 
0560 out:
0561     rcu_read_unlock();
0562     return anon_vma;
0563 }
0564 
0565 void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
0566 {
0567     anon_vma_unlock_read(anon_vma);
0568 }
0569 
0570 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
0571 /*
0572  * Flush TLB entries for recently unmapped pages from remote CPUs. It is
0573  * important if a PTE was dirty when it was unmapped that it's flushed
0574  * before any IO is initiated on the page to prevent lost writes. Similarly,
0575  * it must be flushed before freeing to prevent data leakage.
0576  */
0577 void try_to_unmap_flush(void)
0578 {
0579     struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
0580     int cpu;
0581 
0582     if (!tlb_ubc->flush_required)
0583         return;
0584 
0585     cpu = get_cpu();
0586 
0587     if (cpumask_test_cpu(cpu, &tlb_ubc->cpumask)) {
0588         count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
0589         local_flush_tlb();
0590         trace_tlb_flush(TLB_LOCAL_SHOOTDOWN, TLB_FLUSH_ALL);
0591     }
0592 
0593     if (cpumask_any_but(&tlb_ubc->cpumask, cpu) < nr_cpu_ids)
0594         flush_tlb_others(&tlb_ubc->cpumask, NULL, 0, TLB_FLUSH_ALL);
0595     cpumask_clear(&tlb_ubc->cpumask);
0596     tlb_ubc->flush_required = false;
0597     tlb_ubc->writable = false;
0598     put_cpu();
0599 }
0600 
0601 /* Flush iff there are potentially writable TLB entries that can race with IO */
0602 void try_to_unmap_flush_dirty(void)
0603 {
0604     struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
0605 
0606     if (tlb_ubc->writable)
0607         try_to_unmap_flush();
0608 }
0609 
0610 static void set_tlb_ubc_flush_pending(struct mm_struct *mm,
0611         struct page *page, bool writable)
0612 {
0613     struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
0614 
0615     cpumask_or(&tlb_ubc->cpumask, &tlb_ubc->cpumask, mm_cpumask(mm));
0616     tlb_ubc->flush_required = true;
0617 
0618     /*
0619      * If the PTE was dirty then it's best to assume it's writable. The
0620      * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
0621      * before the page is queued for IO.
0622      */
0623     if (writable)
0624         tlb_ubc->writable = true;
0625 }
0626 
0627 /*
0628  * Returns true if the TLB flush should be deferred to the end of a batch of
0629  * unmap operations to reduce IPIs.
0630  */
0631 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
0632 {
0633     bool should_defer = false;
0634 
0635     if (!(flags & TTU_BATCH_FLUSH))
0636         return false;
0637 
0638     /* If remote CPUs need to be flushed then defer batch the flush */
0639     if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
0640         should_defer = true;
0641     put_cpu();
0642 
0643     return should_defer;
0644 }
0645 #else
0646 static void set_tlb_ubc_flush_pending(struct mm_struct *mm,
0647         struct page *page, bool writable)
0648 {
0649 }
0650 
0651 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
0652 {
0653     return false;
0654 }
0655 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
0656 
0657 /*
0658  * At what user virtual address is page expected in vma?
0659  * Caller should check the page is actually part of the vma.
0660  */
0661 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
0662 {
0663     unsigned long address;
0664     if (PageAnon(page)) {
0665         struct anon_vma *page__anon_vma = page_anon_vma(page);
0666         /*
0667          * Note: swapoff's unuse_vma() is more efficient with this
0668          * check, and needs it to match anon_vma when KSM is active.
0669          */
0670         if (!vma->anon_vma || !page__anon_vma ||
0671             vma->anon_vma->root != page__anon_vma->root)
0672             return -EFAULT;
0673     } else if (page->mapping) {
0674         if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping)
0675             return -EFAULT;
0676     } else
0677         return -EFAULT;
0678     address = __vma_address(page, vma);
0679     if (unlikely(address < vma->vm_start || address >= vma->vm_end))
0680         return -EFAULT;
0681     return address;
0682 }
0683 
0684 pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
0685 {
0686     pgd_t *pgd;
0687     pud_t *pud;
0688     pmd_t *pmd = NULL;
0689     pmd_t pmde;
0690 
0691     pgd = pgd_offset(mm, address);
0692     if (!pgd_present(*pgd))
0693         goto out;
0694 
0695     pud = pud_offset(pgd, address);
0696     if (!pud_present(*pud))
0697         goto out;
0698 
0699     pmd = pmd_offset(pud, address);
0700     /*
0701      * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
0702      * without holding anon_vma lock for write.  So when looking for a
0703      * genuine pmde (in which to find pte), test present and !THP together.
0704      */
0705     pmde = *pmd;
0706     barrier();
0707     if (!pmd_present(pmde) || pmd_trans_huge(pmde))
0708         pmd = NULL;
0709 out:
0710     return pmd;
0711 }
0712 
0713 /*
0714  * Check that @page is mapped at @address into @mm.
0715  *
0716  * If @sync is false, page_check_address may perform a racy check to avoid
0717  * the page table lock when the pte is not present (helpful when reclaiming
0718  * highly shared pages).
0719  *
0720  * On success returns with pte mapped and locked.
0721  */
0722 pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
0723               unsigned long address, spinlock_t **ptlp, int sync)
0724 {
0725     pmd_t *pmd;
0726     pte_t *pte;
0727     spinlock_t *ptl;
0728 
0729     if (unlikely(PageHuge(page))) {
0730         /* when pud is not present, pte will be NULL */
0731         pte = huge_pte_offset(mm, address);
0732         if (!pte)
0733             return NULL;
0734 
0735         ptl = huge_pte_lockptr(page_hstate(page), mm, pte);
0736         goto check;
0737     }
0738 
0739     pmd = mm_find_pmd(mm, address);
0740     if (!pmd)
0741         return NULL;
0742 
0743     pte = pte_offset_map(pmd, address);
0744     /* Make a quick check before getting the lock */
0745     if (!sync && !pte_present(*pte)) {
0746         pte_unmap(pte);
0747         return NULL;
0748     }
0749 
0750     ptl = pte_lockptr(mm, pmd);
0751 check:
0752     spin_lock(ptl);
0753     if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
0754         *ptlp = ptl;
0755         return pte;
0756     }
0757     pte_unmap_unlock(pte, ptl);
0758     return NULL;
0759 }
0760 
0761 /**
0762  * page_mapped_in_vma - check whether a page is really mapped in a VMA
0763  * @page: the page to test
0764  * @vma: the VMA to test
0765  *
0766  * Returns 1 if the page is mapped into the page tables of the VMA, 0
0767  * if the page is not mapped into the page tables of this VMA.  Only
0768  * valid for normal file or anonymous VMAs.
0769  */
0770 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
0771 {
0772     unsigned long address;
0773     pte_t *pte;
0774     spinlock_t *ptl;
0775 
0776     address = __vma_address(page, vma);
0777     if (unlikely(address < vma->vm_start || address >= vma->vm_end))
0778         return 0;
0779     pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
0780     if (!pte)           /* the page is not in this mm */
0781         return 0;
0782     pte_unmap_unlock(pte, ptl);
0783 
0784     return 1;
0785 }
0786 
0787 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
0788 /*
0789  * Check that @page is mapped at @address into @mm. In contrast to
0790  * page_check_address(), this function can handle transparent huge pages.
0791  *
0792  * On success returns true with pte mapped and locked. For PMD-mapped
0793  * transparent huge pages *@ptep is set to NULL.
0794  */
0795 bool page_check_address_transhuge(struct page *page, struct mm_struct *mm,
0796                   unsigned long address, pmd_t **pmdp,
0797                   pte_t **ptep, spinlock_t **ptlp)
0798 {
0799     pgd_t *pgd;
0800     pud_t *pud;
0801     pmd_t *pmd;
0802     pte_t *pte;
0803     spinlock_t *ptl;
0804 
0805     if (unlikely(PageHuge(page))) {
0806         /* when pud is not present, pte will be NULL */
0807         pte = huge_pte_offset(mm, address);
0808         if (!pte)
0809             return false;
0810 
0811         ptl = huge_pte_lockptr(page_hstate(page), mm, pte);
0812         pmd = NULL;
0813         goto check_pte;
0814     }
0815 
0816     pgd = pgd_offset(mm, address);
0817     if (!pgd_present(*pgd))
0818         return false;
0819     pud = pud_offset(pgd, address);
0820     if (!pud_present(*pud))
0821         return false;
0822     pmd = pmd_offset(pud, address);
0823 
0824     if (pmd_trans_huge(*pmd)) {
0825         ptl = pmd_lock(mm, pmd);
0826         if (!pmd_present(*pmd))
0827             goto unlock_pmd;
0828         if (unlikely(!pmd_trans_huge(*pmd))) {
0829             spin_unlock(ptl);
0830             goto map_pte;
0831         }
0832 
0833         if (pmd_page(*pmd) != page)
0834             goto unlock_pmd;
0835 
0836         pte = NULL;
0837         goto found;
0838 unlock_pmd:
0839         spin_unlock(ptl);
0840         return false;
0841     } else {
0842         pmd_t pmde = *pmd;
0843 
0844         barrier();
0845         if (!pmd_present(pmde) || pmd_trans_huge(pmde))
0846             return false;
0847     }
0848 map_pte:
0849     pte = pte_offset_map(pmd, address);
0850     if (!pte_present(*pte)) {
0851         pte_unmap(pte);
0852         return false;
0853     }
0854 
0855     ptl = pte_lockptr(mm, pmd);
0856 check_pte:
0857     spin_lock(ptl);
0858 
0859     if (!pte_present(*pte)) {
0860         pte_unmap_unlock(pte, ptl);
0861         return false;
0862     }
0863 
0864     /* THP can be referenced by any subpage */
0865     if (pte_pfn(*pte) - page_to_pfn(page) >= hpage_nr_pages(page)) {
0866         pte_unmap_unlock(pte, ptl);
0867         return false;
0868     }
0869 found:
0870     *ptep = pte;
0871     *pmdp = pmd;
0872     *ptlp = ptl;
0873     return true;
0874 }
0875 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
0876 
0877 struct page_referenced_arg {
0878     int mapcount;
0879     int referenced;
0880     unsigned long vm_flags;
0881     struct mem_cgroup *memcg;
0882 };
0883 /*
0884  * arg: page_referenced_arg will be passed
0885  */
0886 static int page_referenced_one(struct page *page, struct vm_area_struct *vma,
0887             unsigned long address, void *arg)
0888 {
0889     struct mm_struct *mm = vma->vm_mm;
0890     struct page_referenced_arg *pra = arg;
0891     pmd_t *pmd;
0892     pte_t *pte;
0893     spinlock_t *ptl;
0894     int referenced = 0;
0895 
0896     if (!page_check_address_transhuge(page, mm, address, &pmd, &pte, &ptl))
0897         return SWAP_AGAIN;
0898 
0899     if (vma->vm_flags & VM_LOCKED) {
0900         if (pte)
0901             pte_unmap(pte);
0902         spin_unlock(ptl);
0903         pra->vm_flags |= VM_LOCKED;
0904         return SWAP_FAIL; /* To break the loop */
0905     }
0906 
0907     if (pte) {
0908         if (ptep_clear_flush_young_notify(vma, address, pte)) {
0909             /*
0910              * Don't treat a reference through a sequentially read
0911              * mapping as such.  If the page has been used in
0912              * another mapping, we will catch it; if this other
0913              * mapping is already gone, the unmap path will have
0914              * set PG_referenced or activated the page.
0915              */
0916             if (likely(!(vma->vm_flags & VM_SEQ_READ)))
0917                 referenced++;
0918         }
0919         pte_unmap(pte);
0920     } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
0921         if (pmdp_clear_flush_young_notify(vma, address, pmd))
0922             referenced++;
0923     } else {
0924         /* unexpected pmd-mapped page? */
0925         WARN_ON_ONCE(1);
0926     }
0927     spin_unlock(ptl);
0928 
0929     if (referenced)
0930         clear_page_idle(page);
0931     if (test_and_clear_page_young(page))
0932         referenced++;
0933 
0934     if (referenced) {
0935         pra->referenced++;
0936         pra->vm_flags |= vma->vm_flags;
0937     }
0938 
0939     pra->mapcount--;
0940     if (!pra->mapcount)
0941         return SWAP_SUCCESS; /* To break the loop */
0942 
0943     return SWAP_AGAIN;
0944 }
0945 
0946 static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
0947 {
0948     struct page_referenced_arg *pra = arg;
0949     struct mem_cgroup *memcg = pra->memcg;
0950 
0951     if (!mm_match_cgroup(vma->vm_mm, memcg))
0952         return true;
0953 
0954     return false;
0955 }
0956 
0957 /**
0958  * page_referenced - test if the page was referenced
0959  * @page: the page to test
0960  * @is_locked: caller holds lock on the page
0961  * @memcg: target memory cgroup
0962  * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
0963  *
0964  * Quick test_and_clear_referenced for all mappings to a page,
0965  * returns the number of ptes which referenced the page.
0966  */
0967 int page_referenced(struct page *page,
0968             int is_locked,
0969             struct mem_cgroup *memcg,
0970             unsigned long *vm_flags)
0971 {
0972     int ret;
0973     int we_locked = 0;
0974     struct page_referenced_arg pra = {
0975         .mapcount = total_mapcount(page),
0976         .memcg = memcg,
0977     };
0978     struct rmap_walk_control rwc = {
0979         .rmap_one = page_referenced_one,
0980         .arg = (void *)&pra,
0981         .anon_lock = page_lock_anon_vma_read,
0982     };
0983 
0984     *vm_flags = 0;
0985     if (!page_mapped(page))
0986         return 0;
0987 
0988     if (!page_rmapping(page))
0989         return 0;
0990 
0991     if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
0992         we_locked = trylock_page(page);
0993         if (!we_locked)
0994             return 1;
0995     }
0996 
0997     /*
0998      * If we are reclaiming on behalf of a cgroup, skip
0999      * counting on behalf of references from different
1000      * cgroups
1001      */
1002     if (memcg) {
1003         rwc.invalid_vma = invalid_page_referenced_vma;
1004     }
1005 
1006     ret = rmap_walk(page, &rwc);
1007     *vm_flags = pra.vm_flags;
1008 
1009     if (we_locked)
1010         unlock_page(page);
1011 
1012     return pra.referenced;
1013 }
1014 
1015 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
1016                 unsigned long address, void *arg)
1017 {
1018     struct mm_struct *mm = vma->vm_mm;
1019     pte_t *pte;
1020     spinlock_t *ptl;
1021     int ret = 0;
1022     int *cleaned = arg;
1023 
1024     pte = page_check_address(page, mm, address, &ptl, 1);
1025     if (!pte)
1026         goto out;
1027 
1028     if (pte_dirty(*pte) || pte_write(*pte)) {
1029         pte_t entry;
1030 
1031         flush_cache_page(vma, address, pte_pfn(*pte));
1032         entry = ptep_clear_flush(vma, address, pte);
1033         entry = pte_wrprotect(entry);
1034         entry = pte_mkclean(entry);
1035         set_pte_at(mm, address, pte, entry);
1036         ret = 1;
1037     }
1038 
1039     pte_unmap_unlock(pte, ptl);
1040 
1041     if (ret) {
1042         mmu_notifier_invalidate_page(mm, address);
1043         (*cleaned)++;
1044     }
1045 out:
1046     return SWAP_AGAIN;
1047 }
1048 
1049 static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
1050 {
1051     if (vma->vm_flags & VM_SHARED)
1052         return false;
1053 
1054     return true;
1055 }
1056 
1057 int page_mkclean(struct page *page)
1058 {
1059     int cleaned = 0;
1060     struct address_space *mapping;
1061     struct rmap_walk_control rwc = {
1062         .arg = (void *)&cleaned,
1063         .rmap_one = page_mkclean_one,
1064         .invalid_vma = invalid_mkclean_vma,
1065     };
1066 
1067     BUG_ON(!PageLocked(page));
1068 
1069     if (!page_mapped(page))
1070         return 0;
1071 
1072     mapping = page_mapping(page);
1073     if (!mapping)
1074         return 0;
1075 
1076     rmap_walk(page, &rwc);
1077 
1078     return cleaned;
1079 }
1080 EXPORT_SYMBOL_GPL(page_mkclean);
1081 
1082 /**
1083  * page_move_anon_rmap - move a page to our anon_vma
1084  * @page:   the page to move to our anon_vma
1085  * @vma:    the vma the page belongs to
1086  *
1087  * When a page belongs exclusively to one process after a COW event,
1088  * that page can be moved into the anon_vma that belongs to just that
1089  * process, so the rmap code will not search the parent or sibling
1090  * processes.
1091  */
1092 void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
1093 {
1094     struct anon_vma *anon_vma = vma->anon_vma;
1095 
1096     page = compound_head(page);
1097 
1098     VM_BUG_ON_PAGE(!PageLocked(page), page);
1099     VM_BUG_ON_VMA(!anon_vma, vma);
1100 
1101     anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1102     /*
1103      * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1104      * simultaneously, so a concurrent reader (eg page_referenced()'s
1105      * PageAnon()) will not see one without the other.
1106      */
1107     WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1108 }
1109 
1110 /**
1111  * __page_set_anon_rmap - set up new anonymous rmap
1112  * @page:   Page to add to rmap 
1113  * @vma:    VM area to add page to.
1114  * @address:    User virtual address of the mapping 
1115  * @exclusive:  the page is exclusively owned by the current process
1116  */
1117 static void __page_set_anon_rmap(struct page *page,
1118     struct vm_area_struct *vma, unsigned long address, int exclusive)
1119 {
1120     struct anon_vma *anon_vma = vma->anon_vma;
1121 
1122     BUG_ON(!anon_vma);
1123 
1124     if (PageAnon(page))
1125         return;
1126 
1127     /*
1128      * If the page isn't exclusively mapped into this vma,
1129      * we must use the _oldest_ possible anon_vma for the
1130      * page mapping!
1131      */
1132     if (!exclusive)
1133         anon_vma = anon_vma->root;
1134 
1135     anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1136     page->mapping = (struct address_space *) anon_vma;
1137     page->index = linear_page_index(vma, address);
1138 }
1139 
1140 /**
1141  * __page_check_anon_rmap - sanity check anonymous rmap addition
1142  * @page:   the page to add the mapping to
1143  * @vma:    the vm area in which the mapping is added
1144  * @address:    the user virtual address mapped
1145  */
1146 static void __page_check_anon_rmap(struct page *page,
1147     struct vm_area_struct *vma, unsigned long address)
1148 {
1149 #ifdef CONFIG_DEBUG_VM
1150     /*
1151      * The page's anon-rmap details (mapping and index) are guaranteed to
1152      * be set up correctly at this point.
1153      *
1154      * We have exclusion against page_add_anon_rmap because the caller
1155      * always holds the page locked, except if called from page_dup_rmap,
1156      * in which case the page is already known to be setup.
1157      *
1158      * We have exclusion against page_add_new_anon_rmap because those pages
1159      * are initially only visible via the pagetables, and the pte is locked
1160      * over the call to page_add_new_anon_rmap.
1161      */
1162     BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1163     BUG_ON(page_to_pgoff(page) != linear_page_index(vma, address));
1164 #endif
1165 }
1166 
1167 /**
1168  * page_add_anon_rmap - add pte mapping to an anonymous page
1169  * @page:   the page to add the mapping to
1170  * @vma:    the vm area in which the mapping is added
1171  * @address:    the user virtual address mapped
1172  * @compound:   charge the page as compound or small page
1173  *
1174  * The caller needs to hold the pte lock, and the page must be locked in
1175  * the anon_vma case: to serialize mapping,index checking after setting,
1176  * and to ensure that PageAnon is not being upgraded racily to PageKsm
1177  * (but PageKsm is never downgraded to PageAnon).
1178  */
1179 void page_add_anon_rmap(struct page *page,
1180     struct vm_area_struct *vma, unsigned long address, bool compound)
1181 {
1182     do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
1183 }
1184 
1185 /*
1186  * Special version of the above for do_swap_page, which often runs
1187  * into pages that are exclusively owned by the current process.
1188  * Everybody else should continue to use page_add_anon_rmap above.
1189  */
1190 void do_page_add_anon_rmap(struct page *page,
1191     struct vm_area_struct *vma, unsigned long address, int flags)
1192 {
1193     bool compound = flags & RMAP_COMPOUND;
1194     bool first;
1195 
1196     if (compound) {
1197         atomic_t *mapcount;
1198         VM_BUG_ON_PAGE(!PageLocked(page), page);
1199         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1200         mapcount = compound_mapcount_ptr(page);
1201         first = atomic_inc_and_test(mapcount);
1202     } else {
1203         first = atomic_inc_and_test(&page->_mapcount);
1204     }
1205 
1206     if (first) {
1207         int nr = compound ? hpage_nr_pages(page) : 1;
1208         /*
1209          * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1210          * these counters are not modified in interrupt context, and
1211          * pte lock(a spinlock) is held, which implies preemption
1212          * disabled.
1213          */
1214         if (compound)
1215             __inc_node_page_state(page, NR_ANON_THPS);
1216         __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1217     }
1218     if (unlikely(PageKsm(page)))
1219         return;
1220 
1221     VM_BUG_ON_PAGE(!PageLocked(page), page);
1222 
1223     /* address might be in next vma when migration races vma_adjust */
1224     if (first)
1225         __page_set_anon_rmap(page, vma, address,
1226                 flags & RMAP_EXCLUSIVE);
1227     else
1228         __page_check_anon_rmap(page, vma, address);
1229 }
1230 
1231 /**
1232  * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1233  * @page:   the page to add the mapping to
1234  * @vma:    the vm area in which the mapping is added
1235  * @address:    the user virtual address mapped
1236  * @compound:   charge the page as compound or small page
1237  *
1238  * Same as page_add_anon_rmap but must only be called on *new* pages.
1239  * This means the inc-and-test can be bypassed.
1240  * Page does not have to be locked.
1241  */
1242 void page_add_new_anon_rmap(struct page *page,
1243     struct vm_area_struct *vma, unsigned long address, bool compound)
1244 {
1245     int nr = compound ? hpage_nr_pages(page) : 1;
1246 
1247     VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1248     __SetPageSwapBacked(page);
1249     if (compound) {
1250         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1251         /* increment count (starts at -1) */
1252         atomic_set(compound_mapcount_ptr(page), 0);
1253         __inc_node_page_state(page, NR_ANON_THPS);
1254     } else {
1255         /* Anon THP always mapped first with PMD */
1256         VM_BUG_ON_PAGE(PageTransCompound(page), page);
1257         /* increment count (starts at -1) */
1258         atomic_set(&page->_mapcount, 0);
1259     }
1260     __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1261     __page_set_anon_rmap(page, vma, address, 1);
1262 }
1263 
1264 /**
1265  * page_add_file_rmap - add pte mapping to a file page
1266  * @page: the page to add the mapping to
1267  *
1268  * The caller needs to hold the pte lock.
1269  */
1270 void page_add_file_rmap(struct page *page, bool compound)
1271 {
1272     int i, nr = 1;
1273 
1274     VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1275     lock_page_memcg(page);
1276     if (compound && PageTransHuge(page)) {
1277         for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1278             if (atomic_inc_and_test(&page[i]._mapcount))
1279                 nr++;
1280         }
1281         if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1282             goto out;
1283         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
1284         __inc_node_page_state(page, NR_SHMEM_PMDMAPPED);
1285     } else {
1286         if (PageTransCompound(page) && page_mapping(page)) {
1287             VM_WARN_ON_ONCE(!PageLocked(page));
1288 
1289             SetPageDoubleMap(compound_head(page));
1290             if (PageMlocked(page))
1291                 clear_page_mlock(compound_head(page));
1292         }
1293         if (!atomic_inc_and_test(&page->_mapcount))
1294             goto out;
1295     }
1296     __mod_node_page_state(page_pgdat(page), NR_FILE_MAPPED, nr);
1297     mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED);
1298 out:
1299     unlock_page_memcg(page);
1300 }
1301 
1302 static void page_remove_file_rmap(struct page *page, bool compound)
1303 {
1304     int i, nr = 1;
1305 
1306     VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1307     lock_page_memcg(page);
1308 
1309     /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1310     if (unlikely(PageHuge(page))) {
1311         /* hugetlb pages are always mapped with pmds */
1312         atomic_dec(compound_mapcount_ptr(page));
1313         goto out;
1314     }
1315 
1316     /* page still mapped by someone else? */
1317     if (compound && PageTransHuge(page)) {
1318         for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1319             if (atomic_add_negative(-1, &page[i]._mapcount))
1320                 nr++;
1321         }
1322         if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1323             goto out;
1324         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
1325         __dec_node_page_state(page, NR_SHMEM_PMDMAPPED);
1326     } else {
1327         if (!atomic_add_negative(-1, &page->_mapcount))
1328             goto out;
1329     }
1330 
1331     /*
1332      * We use the irq-unsafe __{inc|mod}_zone_page_state because
1333      * these counters are not modified in interrupt context, and
1334      * pte lock(a spinlock) is held, which implies preemption disabled.
1335      */
1336     __mod_node_page_state(page_pgdat(page), NR_FILE_MAPPED, -nr);
1337     mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED);
1338 
1339     if (unlikely(PageMlocked(page)))
1340         clear_page_mlock(page);
1341 out:
1342     unlock_page_memcg(page);
1343 }
1344 
1345 static void page_remove_anon_compound_rmap(struct page *page)
1346 {
1347     int i, nr;
1348 
1349     if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1350         return;
1351 
1352     /* Hugepages are not counted in NR_ANON_PAGES for now. */
1353     if (unlikely(PageHuge(page)))
1354         return;
1355 
1356     if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1357         return;
1358 
1359     __dec_node_page_state(page, NR_ANON_THPS);
1360 
1361     if (TestClearPageDoubleMap(page)) {
1362         /*
1363          * Subpages can be mapped with PTEs too. Check how many of
1364          * themi are still mapped.
1365          */
1366         for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1367             if (atomic_add_negative(-1, &page[i]._mapcount))
1368                 nr++;
1369         }
1370     } else {
1371         nr = HPAGE_PMD_NR;
1372     }
1373 
1374     if (unlikely(PageMlocked(page)))
1375         clear_page_mlock(page);
1376 
1377     if (nr) {
1378         __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, -nr);
1379         deferred_split_huge_page(page);
1380     }
1381 }
1382 
1383 /**
1384  * page_remove_rmap - take down pte mapping from a page
1385  * @page:   page to remove mapping from
1386  * @compound:   uncharge the page as compound or small page
1387  *
1388  * The caller needs to hold the pte lock.
1389  */
1390 void page_remove_rmap(struct page *page, bool compound)
1391 {
1392     if (!PageAnon(page))
1393         return page_remove_file_rmap(page, compound);
1394 
1395     if (compound)
1396         return page_remove_anon_compound_rmap(page);
1397 
1398     /* page still mapped by someone else? */
1399     if (!atomic_add_negative(-1, &page->_mapcount))
1400         return;
1401 
1402     /*
1403      * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1404      * these counters are not modified in interrupt context, and
1405      * pte lock(a spinlock) is held, which implies preemption disabled.
1406      */
1407     __dec_node_page_state(page, NR_ANON_MAPPED);
1408 
1409     if (unlikely(PageMlocked(page)))
1410         clear_page_mlock(page);
1411 
1412     if (PageTransCompound(page))
1413         deferred_split_huge_page(compound_head(page));
1414 
1415     /*
1416      * It would be tidy to reset the PageAnon mapping here,
1417      * but that might overwrite a racing page_add_anon_rmap
1418      * which increments mapcount after us but sets mapping
1419      * before us: so leave the reset to free_hot_cold_page,
1420      * and remember that it's only reliable while mapped.
1421      * Leaving it set also helps swapoff to reinstate ptes
1422      * faster for those pages still in swapcache.
1423      */
1424 }
1425 
1426 struct rmap_private {
1427     enum ttu_flags flags;
1428     int lazyfreed;
1429 };
1430 
1431 /*
1432  * @arg: enum ttu_flags will be passed to this argument
1433  */
1434 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1435              unsigned long address, void *arg)
1436 {
1437     struct mm_struct *mm = vma->vm_mm;
1438     pte_t *pte;
1439     pte_t pteval;
1440     spinlock_t *ptl;
1441     int ret = SWAP_AGAIN;
1442     struct rmap_private *rp = arg;
1443     enum ttu_flags flags = rp->flags;
1444 
1445     /* munlock has nothing to gain from examining un-locked vmas */
1446     if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
1447         goto out;
1448 
1449     if (flags & TTU_SPLIT_HUGE_PMD) {
1450         split_huge_pmd_address(vma, address,
1451                 flags & TTU_MIGRATION, page);
1452         /* check if we have anything to do after split */
1453         if (page_mapcount(page) == 0)
1454             goto out;
1455     }
1456 
1457     pte = page_check_address(page, mm, address, &ptl,
1458                  PageTransCompound(page));
1459     if (!pte)
1460         goto out;
1461 
1462     /*
1463      * If the page is mlock()d, we cannot swap it out.
1464      * If it's recently referenced (perhaps page_referenced
1465      * skipped over this mm) then we should reactivate it.
1466      */
1467     if (!(flags & TTU_IGNORE_MLOCK)) {
1468         if (vma->vm_flags & VM_LOCKED) {
1469             /* PTE-mapped THP are never mlocked */
1470             if (!PageTransCompound(page)) {
1471                 /*
1472                  * Holding pte lock, we do *not* need
1473                  * mmap_sem here
1474                  */
1475                 mlock_vma_page(page);
1476             }
1477             ret = SWAP_MLOCK;
1478             goto out_unmap;
1479         }
1480         if (flags & TTU_MUNLOCK)
1481             goto out_unmap;
1482     }
1483     if (!(flags & TTU_IGNORE_ACCESS)) {
1484         if (ptep_clear_flush_young_notify(vma, address, pte)) {
1485             ret = SWAP_FAIL;
1486             goto out_unmap;
1487         }
1488     }
1489 
1490     /* Nuke the page table entry. */
1491     flush_cache_page(vma, address, page_to_pfn(page));
1492     if (should_defer_flush(mm, flags)) {
1493         /*
1494          * We clear the PTE but do not flush so potentially a remote
1495          * CPU could still be writing to the page. If the entry was
1496          * previously clean then the architecture must guarantee that
1497          * a clear->dirty transition on a cached TLB entry is written
1498          * through and traps if the PTE is unmapped.
1499          */
1500         pteval = ptep_get_and_clear(mm, address, pte);
1501 
1502         set_tlb_ubc_flush_pending(mm, page, pte_dirty(pteval));
1503     } else {
1504         pteval = ptep_clear_flush(vma, address, pte);
1505     }
1506 
1507     /* Move the dirty bit to the physical page now the pte is gone. */
1508     if (pte_dirty(pteval))
1509         set_page_dirty(page);
1510 
1511     /* Update high watermark before we lower rss */
1512     update_hiwater_rss(mm);
1513 
1514     if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1515         if (PageHuge(page)) {
1516             hugetlb_count_sub(1 << compound_order(page), mm);
1517         } else {
1518             dec_mm_counter(mm, mm_counter(page));
1519         }
1520         set_pte_at(mm, address, pte,
1521                swp_entry_to_pte(make_hwpoison_entry(page)));
1522     } else if (pte_unused(pteval)) {
1523         /*
1524          * The guest indicated that the page content is of no
1525          * interest anymore. Simply discard the pte, vmscan
1526          * will take care of the rest.
1527          */
1528         dec_mm_counter(mm, mm_counter(page));
1529     } else if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION)) {
1530         swp_entry_t entry;
1531         pte_t swp_pte;
1532         /*
1533          * Store the pfn of the page in a special migration
1534          * pte. do_swap_page() will wait until the migration
1535          * pte is removed and then restart fault handling.
1536          */
1537         entry = make_migration_entry(page, pte_write(pteval));
1538         swp_pte = swp_entry_to_pte(entry);
1539         if (pte_soft_dirty(pteval))
1540             swp_pte = pte_swp_mksoft_dirty(swp_pte);
1541         set_pte_at(mm, address, pte, swp_pte);
1542     } else if (PageAnon(page)) {
1543         swp_entry_t entry = { .val = page_private(page) };
1544         pte_t swp_pte;
1545         /*
1546          * Store the swap location in the pte.
1547          * See handle_pte_fault() ...
1548          */
1549         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
1550 
1551         if (!PageDirty(page) && (flags & TTU_LZFREE)) {
1552             /* It's a freeable page by MADV_FREE */
1553             dec_mm_counter(mm, MM_ANONPAGES);
1554             rp->lazyfreed++;
1555             goto discard;
1556         }
1557 
1558         if (swap_duplicate(entry) < 0) {
1559             set_pte_at(mm, address, pte, pteval);
1560             ret = SWAP_FAIL;
1561             goto out_unmap;
1562         }
1563         if (list_empty(&mm->mmlist)) {
1564             spin_lock(&mmlist_lock);
1565             if (list_empty(&mm->mmlist))
1566                 list_add(&mm->mmlist, &init_mm.mmlist);
1567             spin_unlock(&mmlist_lock);
1568         }
1569         dec_mm_counter(mm, MM_ANONPAGES);
1570         inc_mm_counter(mm, MM_SWAPENTS);
1571         swp_pte = swp_entry_to_pte(entry);
1572         if (pte_soft_dirty(pteval))
1573             swp_pte = pte_swp_mksoft_dirty(swp_pte);
1574         set_pte_at(mm, address, pte, swp_pte);
1575     } else
1576         dec_mm_counter(mm, mm_counter_file(page));
1577 
1578 discard:
1579     page_remove_rmap(page, PageHuge(page));
1580     put_page(page);
1581 
1582 out_unmap:
1583     pte_unmap_unlock(pte, ptl);
1584     if (ret != SWAP_FAIL && ret != SWAP_MLOCK && !(flags & TTU_MUNLOCK))
1585         mmu_notifier_invalidate_page(mm, address);
1586 out:
1587     return ret;
1588 }
1589 
1590 bool is_vma_temporary_stack(struct vm_area_struct *vma)
1591 {
1592     int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1593 
1594     if (!maybe_stack)
1595         return false;
1596 
1597     if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1598                         VM_STACK_INCOMPLETE_SETUP)
1599         return true;
1600 
1601     return false;
1602 }
1603 
1604 static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1605 {
1606     return is_vma_temporary_stack(vma);
1607 }
1608 
1609 static int page_mapcount_is_zero(struct page *page)
1610 {
1611     return !page_mapcount(page);
1612 }
1613 
1614 /**
1615  * try_to_unmap - try to remove all page table mappings to a page
1616  * @page: the page to get unmapped
1617  * @flags: action and flags
1618  *
1619  * Tries to remove all the page table entries which are mapping this
1620  * page, used in the pageout path.  Caller must hold the page lock.
1621  * Return values are:
1622  *
1623  * SWAP_SUCCESS - we succeeded in removing all mappings
1624  * SWAP_AGAIN   - we missed a mapping, try again later
1625  * SWAP_FAIL    - the page is unswappable
1626  * SWAP_MLOCK   - page is mlocked.
1627  */
1628 int try_to_unmap(struct page *page, enum ttu_flags flags)
1629 {
1630     int ret;
1631     struct rmap_private rp = {
1632         .flags = flags,
1633         .lazyfreed = 0,
1634     };
1635 
1636     struct rmap_walk_control rwc = {
1637         .rmap_one = try_to_unmap_one,
1638         .arg = &rp,
1639         .done = page_mapcount_is_zero,
1640         .anon_lock = page_lock_anon_vma_read,
1641     };
1642 
1643     /*
1644      * During exec, a temporary VMA is setup and later moved.
1645      * The VMA is moved under the anon_vma lock but not the
1646      * page tables leading to a race where migration cannot
1647      * find the migration ptes. Rather than increasing the
1648      * locking requirements of exec(), migration skips
1649      * temporary VMAs until after exec() completes.
1650      */
1651     if ((flags & TTU_MIGRATION) && !PageKsm(page) && PageAnon(page))
1652         rwc.invalid_vma = invalid_migration_vma;
1653 
1654     if (flags & TTU_RMAP_LOCKED)
1655         ret = rmap_walk_locked(page, &rwc);
1656     else
1657         ret = rmap_walk(page, &rwc);
1658 
1659     if (ret != SWAP_MLOCK && !page_mapcount(page)) {
1660         ret = SWAP_SUCCESS;
1661         if (rp.lazyfreed && !PageDirty(page))
1662             ret = SWAP_LZFREE;
1663     }
1664     return ret;
1665 }
1666 
1667 static int page_not_mapped(struct page *page)
1668 {
1669     return !page_mapped(page);
1670 };
1671 
1672 /**
1673  * try_to_munlock - try to munlock a page
1674  * @page: the page to be munlocked
1675  *
1676  * Called from munlock code.  Checks all of the VMAs mapping the page
1677  * to make sure nobody else has this page mlocked. The page will be
1678  * returned with PG_mlocked cleared if no other vmas have it mlocked.
1679  *
1680  * Return values are:
1681  *
1682  * SWAP_AGAIN   - no vma is holding page mlocked, or,
1683  * SWAP_AGAIN   - page mapped in mlocked vma -- couldn't acquire mmap sem
1684  * SWAP_FAIL    - page cannot be located at present
1685  * SWAP_MLOCK   - page is now mlocked.
1686  */
1687 int try_to_munlock(struct page *page)
1688 {
1689     int ret;
1690     struct rmap_private rp = {
1691         .flags = TTU_MUNLOCK,
1692         .lazyfreed = 0,
1693     };
1694 
1695     struct rmap_walk_control rwc = {
1696         .rmap_one = try_to_unmap_one,
1697         .arg = &rp,
1698         .done = page_not_mapped,
1699         .anon_lock = page_lock_anon_vma_read,
1700 
1701     };
1702 
1703     VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1704 
1705     ret = rmap_walk(page, &rwc);
1706     return ret;
1707 }
1708 
1709 void __put_anon_vma(struct anon_vma *anon_vma)
1710 {
1711     struct anon_vma *root = anon_vma->root;
1712 
1713     anon_vma_free(anon_vma);
1714     if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1715         anon_vma_free(root);
1716 }
1717 
1718 static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1719                     struct rmap_walk_control *rwc)
1720 {
1721     struct anon_vma *anon_vma;
1722 
1723     if (rwc->anon_lock)
1724         return rwc->anon_lock(page);
1725 
1726     /*
1727      * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1728      * because that depends on page_mapped(); but not all its usages
1729      * are holding mmap_sem. Users without mmap_sem are required to
1730      * take a reference count to prevent the anon_vma disappearing
1731      */
1732     anon_vma = page_anon_vma(page);
1733     if (!anon_vma)
1734         return NULL;
1735 
1736     anon_vma_lock_read(anon_vma);
1737     return anon_vma;
1738 }
1739 
1740 /*
1741  * rmap_walk_anon - do something to anonymous page using the object-based
1742  * rmap method
1743  * @page: the page to be handled
1744  * @rwc: control variable according to each walk type
1745  *
1746  * Find all the mappings of a page using the mapping pointer and the vma chains
1747  * contained in the anon_vma struct it points to.
1748  *
1749  * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1750  * where the page was found will be held for write.  So, we won't recheck
1751  * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1752  * LOCKED.
1753  */
1754 static int rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
1755         bool locked)
1756 {
1757     struct anon_vma *anon_vma;
1758     pgoff_t pgoff;
1759     struct anon_vma_chain *avc;
1760     int ret = SWAP_AGAIN;
1761 
1762     if (locked) {
1763         anon_vma = page_anon_vma(page);
1764         /* anon_vma disappear under us? */
1765         VM_BUG_ON_PAGE(!anon_vma, page);
1766     } else {
1767         anon_vma = rmap_walk_anon_lock(page, rwc);
1768     }
1769     if (!anon_vma)
1770         return ret;
1771 
1772     pgoff = page_to_pgoff(page);
1773     anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1774         struct vm_area_struct *vma = avc->vma;
1775         unsigned long address = vma_address(page, vma);
1776 
1777         cond_resched();
1778 
1779         if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1780             continue;
1781 
1782         ret = rwc->rmap_one(page, vma, address, rwc->arg);
1783         if (ret != SWAP_AGAIN)
1784             break;
1785         if (rwc->done && rwc->done(page))
1786             break;
1787     }
1788 
1789     if (!locked)
1790         anon_vma_unlock_read(anon_vma);
1791     return ret;
1792 }
1793 
1794 /*
1795  * rmap_walk_file - do something to file page using the object-based rmap method
1796  * @page: the page to be handled
1797  * @rwc: control variable according to each walk type
1798  *
1799  * Find all the mappings of a page using the mapping pointer and the vma chains
1800  * contained in the address_space struct it points to.
1801  *
1802  * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1803  * where the page was found will be held for write.  So, we won't recheck
1804  * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1805  * LOCKED.
1806  */
1807 static int rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
1808         bool locked)
1809 {
1810     struct address_space *mapping = page_mapping(page);
1811     pgoff_t pgoff;
1812     struct vm_area_struct *vma;
1813     int ret = SWAP_AGAIN;
1814 
1815     /*
1816      * The page lock not only makes sure that page->mapping cannot
1817      * suddenly be NULLified by truncation, it makes sure that the
1818      * structure at mapping cannot be freed and reused yet,
1819      * so we can safely take mapping->i_mmap_rwsem.
1820      */
1821     VM_BUG_ON_PAGE(!PageLocked(page), page);
1822 
1823     if (!mapping)
1824         return ret;
1825 
1826     pgoff = page_to_pgoff(page);
1827     if (!locked)
1828         i_mmap_lock_read(mapping);
1829     vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1830         unsigned long address = vma_address(page, vma);
1831 
1832         cond_resched();
1833 
1834         if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1835             continue;
1836 
1837         ret = rwc->rmap_one(page, vma, address, rwc->arg);
1838         if (ret != SWAP_AGAIN)
1839             goto done;
1840         if (rwc->done && rwc->done(page))
1841             goto done;
1842     }
1843 
1844 done:
1845     if (!locked)
1846         i_mmap_unlock_read(mapping);
1847     return ret;
1848 }
1849 
1850 int rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1851 {
1852     if (unlikely(PageKsm(page)))
1853         return rmap_walk_ksm(page, rwc);
1854     else if (PageAnon(page))
1855         return rmap_walk_anon(page, rwc, false);
1856     else
1857         return rmap_walk_file(page, rwc, false);
1858 }
1859 
1860 /* Like rmap_walk, but caller holds relevant rmap lock */
1861 int rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
1862 {
1863     /* no ksm support for now */
1864     VM_BUG_ON_PAGE(PageKsm(page), page);
1865     if (PageAnon(page))
1866         return rmap_walk_anon(page, rwc, true);
1867     else
1868         return rmap_walk_file(page, rwc, true);
1869 }
1870 
1871 #ifdef CONFIG_HUGETLB_PAGE
1872 /*
1873  * The following three functions are for anonymous (private mapped) hugepages.
1874  * Unlike common anonymous pages, anonymous hugepages have no accounting code
1875  * and no lru code, because we handle hugepages differently from common pages.
1876  */
1877 static void __hugepage_set_anon_rmap(struct page *page,
1878     struct vm_area_struct *vma, unsigned long address, int exclusive)
1879 {
1880     struct anon_vma *anon_vma = vma->anon_vma;
1881 
1882     BUG_ON(!anon_vma);
1883 
1884     if (PageAnon(page))
1885         return;
1886     if (!exclusive)
1887         anon_vma = anon_vma->root;
1888 
1889     anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1890     page->mapping = (struct address_space *) anon_vma;
1891     page->index = linear_page_index(vma, address);
1892 }
1893 
1894 void hugepage_add_anon_rmap(struct page *page,
1895                 struct vm_area_struct *vma, unsigned long address)
1896 {
1897     struct anon_vma *anon_vma = vma->anon_vma;
1898     int first;
1899 
1900     BUG_ON(!PageLocked(page));
1901     BUG_ON(!anon_vma);
1902     /* address might be in next vma when migration races vma_adjust */
1903     first = atomic_inc_and_test(compound_mapcount_ptr(page));
1904     if (first)
1905         __hugepage_set_anon_rmap(page, vma, address, 0);
1906 }
1907 
1908 void hugepage_add_new_anon_rmap(struct page *page,
1909             struct vm_area_struct *vma, unsigned long address)
1910 {
1911     BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1912     atomic_set(compound_mapcount_ptr(page), 0);
1913     __hugepage_set_anon_rmap(page, vma, address, 1);
1914 }
1915 #endif /* CONFIG_HUGETLB_PAGE */