mm: bugfix: set current->reclaim_state to NULL while returning from kswapd()
[linux-2.6.git] / mm / rmap.c
index dd43373..bfca52c 100644 (file)
--- a/mm/rmap.c
+++ b/mm/rmap.c
  * Lock ordering in mm:
  *
  * inode->i_mutex      (while writing or truncating, not reading or faulting)
- *   inode->i_alloc_sem (vmtruncate_range)
  *   mm->mmap_sem
  *     page->flags PG_locked (lock_page)
- *       mapping->i_mmap_lock
- *         anon_vma->lock
+ *       mapping->i_mmap_mutex
+ *         anon_vma->mutex
  *           mm->page_table_lock or pte_lock
  *             zone->lru_lock (in mark_page_accessed, isolate_lru_page)
  *             swap_lock (in swap_duplicate, swap_info_get)
  *               mmlist_lock (in mmput, drain_mmlist and others)
  *               mapping->private_lock (in __set_page_dirty_buffers)
- *               inode_lock (in set_page_dirty's __mark_inode_dirty)
+ *               inode->i_lock (in set_page_dirty's __mark_inode_dirty)
+ *               bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
  *                 sb_lock (within inode_lock in fs/fs-writeback.c)
  *                 mapping->tree_lock (widely used, in set_page_dirty,
  *                           in arch-dependent flush_dcache_mmap_lock,
- *                           within inode_lock in __sync_single_inode)
+ *                           within bdi.wb->list_lock in __sync_single_inode)
  *
- * (code doesn't rely on that order so it could be switched around)
- * ->tasklist_lock
- *   anon_vma->lock      (memory_failure, collect_procs_anon)
+ * anon_vma->mutex,mapping->i_mutex      (memory_failure, collect_procs_anon)
+ *   ->tasklist_lock
  *     pte map lock
  */
 
 #include <linux/swapops.h>
 #include <linux/slab.h>
 #include <linux/init.h>
+#include <linux/ksm.h>
 #include <linux/rmap.h>
 #include <linux/rcupdate.h>
-#include <linux/module.h>
+#include <linux/export.h>
 #include <linux/memcontrol.h>
 #include <linux/mmu_notifier.h>
 #include <linux/migrate.h>
+#include <linux/hugetlb.h>
+#include <linux/backing-dev.h>
 
 #include <asm/tlbflush.h>
 
 #include "internal.h"
 
 static struct kmem_cache *anon_vma_cachep;
+static struct kmem_cache *anon_vma_chain_cachep;
 
 static inline struct anon_vma *anon_vma_alloc(void)
 {
-       return kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
+       struct anon_vma *anon_vma;
+
+       anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
+       if (anon_vma) {
+               atomic_set(&anon_vma->refcount, 1);
+               /*
+                * Initialise the anon_vma root to point to itself. If called
+                * from fork, the root will be reset to the parents anon_vma.
+                */
+               anon_vma->root = anon_vma;
+       }
+
+       return anon_vma;
 }
 
 static inline void anon_vma_free(struct anon_vma *anon_vma)
 {
+       VM_BUG_ON(atomic_read(&anon_vma->refcount));
+
+       /*
+        * Synchronize against page_lock_anon_vma() such that
+        * we can safely hold the lock without the anon_vma getting
+        * freed.
+        *
+        * Relies on the full mb implied by the atomic_dec_and_test() from
+        * put_anon_vma() against the acquire barrier implied by
+        * mutex_trylock() from page_lock_anon_vma(). This orders:
+        *
+        * page_lock_anon_vma()         VS      put_anon_vma()
+        *   mutex_trylock()                      atomic_dec_and_test()
+        *   LOCK                                 MB
+        *   atomic_read()                        mutex_is_locked()
+        *
+        * LOCK should suffice since the actual taking of the lock must
+        * happen _before_ what follows.
+        */
+       if (mutex_is_locked(&anon_vma->root->mutex)) {
+               anon_vma_lock(anon_vma);
+               anon_vma_unlock(anon_vma);
+       }
+
        kmem_cache_free(anon_vma_cachep, anon_vma);
 }
 
+static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
+{
+       return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
+}
+
+static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
+{
+       kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
+}
+
+static void anon_vma_chain_link(struct vm_area_struct *vma,
+                               struct anon_vma_chain *avc,
+                               struct anon_vma *anon_vma)
+{
+       avc->vma = vma;
+       avc->anon_vma = anon_vma;
+       list_add(&avc->same_vma, &vma->anon_vma_chain);
+
+       /*
+        * It's critical to add new vmas to the tail of the anon_vma,
+        * see comment in huge_memory.c:__split_huge_page().
+        */
+       list_add_tail(&avc->same_anon_vma, &anon_vma->head);
+}
+
 /**
  * anon_vma_prepare - attach an anon_vma to a memory region
  * @vma: the memory region in question
@@ -81,7 +145,7 @@ static inline void anon_vma_free(struct anon_vma *anon_vma)
  * anonymous pages mapped into it with that anon_vma.
  *
  * The common case will be that we already have one, but if
- * if not we either need to find an adjacent mapping that we
+ * not we either need to find an adjacent mapping that we
  * can re-use the anon_vma from (very common when the only
  * reason for splitting a vma has been mprotect()), or we
  * allocate a new one.
@@ -102,87 +166,257 @@ static inline void anon_vma_free(struct anon_vma *anon_vma)
 int anon_vma_prepare(struct vm_area_struct *vma)
 {
        struct anon_vma *anon_vma = vma->anon_vma;
+       struct anon_vma_chain *avc;
 
        might_sleep();
        if (unlikely(!anon_vma)) {
                struct mm_struct *mm = vma->vm_mm;
                struct anon_vma *allocated;
 
+               avc = anon_vma_chain_alloc(GFP_KERNEL);
+               if (!avc)
+                       goto out_enomem;
+
                anon_vma = find_mergeable_anon_vma(vma);
                allocated = NULL;
                if (!anon_vma) {
                        anon_vma = anon_vma_alloc();
                        if (unlikely(!anon_vma))
-                               return -ENOMEM;
+                               goto out_enomem_free_avc;
                        allocated = anon_vma;
                }
-               spin_lock(&anon_vma->lock);
 
+               anon_vma_lock(anon_vma);
                /* page_table_lock to protect against threads */
                spin_lock(&mm->page_table_lock);
                if (likely(!vma->anon_vma)) {
                        vma->anon_vma = anon_vma;
-                       list_add_tail(&vma->anon_vma_node, &anon_vma->head);
+                       anon_vma_chain_link(vma, avc, anon_vma);
                        allocated = NULL;
+                       avc = NULL;
                }
                spin_unlock(&mm->page_table_lock);
+               anon_vma_unlock(anon_vma);
 
-               spin_unlock(&anon_vma->lock);
                if (unlikely(allocated))
-                       anon_vma_free(allocated);
+                       put_anon_vma(allocated);
+               if (unlikely(avc))
+                       anon_vma_chain_free(avc);
        }
        return 0;
+
+ out_enomem_free_avc:
+       anon_vma_chain_free(avc);
+ out_enomem:
+       return -ENOMEM;
 }
 
-void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
+/*
+ * This is a useful helper function for locking the anon_vma root as
+ * we traverse the vma->anon_vma_chain, looping over anon_vma's that
+ * have the same vma.
+ *
+ * Such anon_vma's should have the same root, so you'd expect to see
+ * just a single mutex_lock for the whole traversal.
+ */
+static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
 {
-       BUG_ON(vma->anon_vma != next->anon_vma);
-       list_del(&next->anon_vma_node);
+       struct anon_vma *new_root = anon_vma->root;
+       if (new_root != root) {
+               if (WARN_ON_ONCE(root))
+                       mutex_unlock(&root->mutex);
+               root = new_root;
+               mutex_lock(&root->mutex);
+       }
+       return root;
 }
 
-void __anon_vma_link(struct vm_area_struct *vma)
+static inline void unlock_anon_vma_root(struct anon_vma *root)
 {
-       struct anon_vma *anon_vma = vma->anon_vma;
-
-       if (anon_vma)
-               list_add_tail(&vma->anon_vma_node, &anon_vma->head);
+       if (root)
+               mutex_unlock(&root->mutex);
 }
 
-void anon_vma_link(struct vm_area_struct *vma)
+/*
+ * Attach the anon_vmas from src to dst.
+ * Returns 0 on success, -ENOMEM on failure.
+ */
+int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
 {
-       struct anon_vma *anon_vma = vma->anon_vma;
+       struct anon_vma_chain *avc, *pavc;
+       struct anon_vma *root = NULL;
+
+       list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
+               struct anon_vma *anon_vma;
+
+               avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
+               if (unlikely(!avc)) {
+                       unlock_anon_vma_root(root);
+                       root = NULL;
+                       avc = anon_vma_chain_alloc(GFP_KERNEL);
+                       if (!avc)
+                               goto enomem_failure;
+               }
+               anon_vma = pavc->anon_vma;
+               root = lock_anon_vma_root(root, anon_vma);
+               anon_vma_chain_link(dst, avc, anon_vma);
+       }
+       unlock_anon_vma_root(root);
+       return 0;
 
-       if (anon_vma) {
-               spin_lock(&anon_vma->lock);
-               list_add_tail(&vma->anon_vma_node, &anon_vma->head);
-               spin_unlock(&anon_vma->lock);
+ enomem_failure:
+       unlink_anon_vmas(dst);
+       return -ENOMEM;
+}
+
+/*
+ * Some rmap walk that needs to find all ptes/hugepmds without false
+ * negatives (like migrate and split_huge_page) running concurrent
+ * with operations that copy or move pagetables (like mremap() and
+ * fork()) to be safe. They depend on the anon_vma "same_anon_vma"
+ * list to be in a certain order: the dst_vma must be placed after the
+ * src_vma in the list. This is always guaranteed by fork() but
+ * mremap() needs to call this function to enforce it in case the
+ * dst_vma isn't newly allocated and chained with the anon_vma_clone()
+ * function but just an extension of a pre-existing vma through
+ * vma_merge.
+ *
+ * NOTE: the same_anon_vma list can still be changed by other
+ * processes while mremap runs because mremap doesn't hold the
+ * anon_vma mutex to prevent modifications to the list while it
+ * runs. All we need to enforce is that the relative order of this
+ * process vmas isn't changing (we don't care about other vmas
+ * order). Each vma corresponds to an anon_vma_chain structure so
+ * there's no risk that other processes calling anon_vma_moveto_tail()
+ * and changing the same_anon_vma list under mremap() will screw with
+ * the relative order of this process vmas in the list, because we
+ * they can't alter the order of any vma that belongs to this
+ * process. And there can't be another anon_vma_moveto_tail() running
+ * concurrently with mremap() coming from this process because we hold
+ * the mmap_sem for the whole mremap(). fork() ordering dependency
+ * also shouldn't be affected because fork() only cares that the
+ * parent vmas are placed in the list before the child vmas and
+ * anon_vma_moveto_tail() won't reorder vmas from either the fork()
+ * parent or child.
+ */
+void anon_vma_moveto_tail(struct vm_area_struct *dst)
+{
+       struct anon_vma_chain *pavc;
+       struct anon_vma *root = NULL;
+
+       list_for_each_entry_reverse(pavc, &dst->anon_vma_chain, same_vma) {
+               struct anon_vma *anon_vma = pavc->anon_vma;
+               VM_BUG_ON(pavc->vma != dst);
+               root = lock_anon_vma_root(root, anon_vma);
+               list_del(&pavc->same_anon_vma);
+               list_add_tail(&pavc->same_anon_vma, &anon_vma->head);
        }
+       unlock_anon_vma_root(root);
 }
 
-void anon_vma_unlink(struct vm_area_struct *vma)
+/*
+ * Attach vma to its own anon_vma, as well as to the anon_vmas that
+ * the corresponding VMA in the parent process is attached to.
+ * Returns 0 on success, non-zero on failure.
+ */
+int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
 {
-       struct anon_vma *anon_vma = vma->anon_vma;
-       int empty;
+       struct anon_vma_chain *avc;
+       struct anon_vma *anon_vma;
+
+       /* Don't bother if the parent process has no anon_vma here. */
+       if (!pvma->anon_vma)
+               return 0;
 
+       /*
+        * First, attach the new VMA to the parent VMA's anon_vmas,
+        * so rmap can find non-COWed pages in child processes.
+        */
+       if (anon_vma_clone(vma, pvma))
+               return -ENOMEM;
+
+       /* Then add our own anon_vma. */
+       anon_vma = anon_vma_alloc();
        if (!anon_vma)
-               return;
+               goto out_error;
+       avc = anon_vma_chain_alloc(GFP_KERNEL);
+       if (!avc)
+               goto out_error_free_anon_vma;
+
+       /*
+        * The root anon_vma's spinlock is the lock actually used when we
+        * lock any of the anon_vmas in this anon_vma tree.
+        */
+       anon_vma->root = pvma->anon_vma->root;
+       /*
+        * With refcounts, an anon_vma can stay around longer than the
+        * process it belongs to. The root anon_vma needs to be pinned until
+        * this anon_vma is freed, because the lock lives in the root.
+        */
+       get_anon_vma(anon_vma->root);
+       /* Mark this anon_vma as the one where our new (COWed) pages go. */
+       vma->anon_vma = anon_vma;
+       anon_vma_lock(anon_vma);
+       anon_vma_chain_link(vma, avc, anon_vma);
+       anon_vma_unlock(anon_vma);
+
+       return 0;
+
+ out_error_free_anon_vma:
+       put_anon_vma(anon_vma);
+ out_error:
+       unlink_anon_vmas(vma);
+       return -ENOMEM;
+}
+
+void unlink_anon_vmas(struct vm_area_struct *vma)
+{
+       struct anon_vma_chain *avc, *next;
+       struct anon_vma *root = NULL;
+
+       /*
+        * Unlink each anon_vma chained to the VMA.  This list is ordered
+        * from newest to oldest, ensuring the root anon_vma gets freed last.
+        */
+       list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
+               struct anon_vma *anon_vma = avc->anon_vma;
 
-       spin_lock(&anon_vma->lock);
-       list_del(&vma->anon_vma_node);
+               root = lock_anon_vma_root(root, anon_vma);
+               list_del(&avc->same_anon_vma);
 
-       /* We must garbage collect the anon_vma if it's empty */
-       empty = list_empty(&anon_vma->head);
-       spin_unlock(&anon_vma->lock);
+               /*
+                * Leave empty anon_vmas on the list - we'll need
+                * to free them outside the lock.
+                */
+               if (list_empty(&anon_vma->head))
+                       continue;
+
+               list_del(&avc->same_vma);
+               anon_vma_chain_free(avc);
+       }
+       unlock_anon_vma_root(root);
+
+       /*
+        * Iterate the list once more, it now only contains empty and unlinked
+        * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
+        * needing to acquire the anon_vma->root->mutex.
+        */
+       list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
+               struct anon_vma *anon_vma = avc->anon_vma;
+
+               put_anon_vma(anon_vma);
 
-       if (empty)
-               anon_vma_free(anon_vma);
+               list_del(&avc->same_vma);
+               anon_vma_chain_free(avc);
+       }
 }
 
 static void anon_vma_ctor(void *data)
 {
        struct anon_vma *anon_vma = data;
 
-       spin_lock_init(&anon_vma->lock);
+       mutex_init(&anon_vma->mutex);
+       atomic_set(&anon_vma->refcount, 0);
        INIT_LIST_HEAD(&anon_vma->head);
 }
 
@@ -190,36 +424,139 @@ void __init anon_vma_init(void)
 {
        anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
                        0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
+       anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
 }
 
 /*
- * Getting a lock on a stable anon_vma from a page off the LRU is
- * tricky: page_lock_anon_vma rely on RCU to guard against the races.
+ * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
+ *
+ * Since there is no serialization what so ever against page_remove_rmap()
+ * the best this function can do is return a locked anon_vma that might
+ * have been relevant to this page.
+ *
+ * The page might have been remapped to a different anon_vma or the anon_vma
+ * returned may already be freed (and even reused).
+ *
+ * In case it was remapped to a different anon_vma, the new anon_vma will be a
+ * child of the old anon_vma, and the anon_vma lifetime rules will therefore
+ * ensure that any anon_vma obtained from the page will still be valid for as
+ * long as we observe page_mapped() [ hence all those page_mapped() tests ].
+ *
+ * All users of this function must be very careful when walking the anon_vma
+ * chain and verify that the page in question is indeed mapped in it
+ * [ something equivalent to page_mapped_in_vma() ].
+ *
+ * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
+ * that the anon_vma pointer from page->mapping is valid if there is a
+ * mapcount, we can dereference the anon_vma after observing those.
+ */
+struct anon_vma *page_get_anon_vma(struct page *page)
+{
+       struct anon_vma *anon_vma = NULL;
+       unsigned long anon_mapping;
+
+       rcu_read_lock();
+       anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
+       if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
+               goto out;
+       if (!page_mapped(page))
+               goto out;
+
+       anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
+       if (!atomic_inc_not_zero(&anon_vma->refcount)) {
+               anon_vma = NULL;
+               goto out;
+       }
+
+       /*
+        * If this page is still mapped, then its anon_vma cannot have been
+        * freed.  But if it has been unmapped, we have no security against the
+        * anon_vma structure being freed and reused (for another anon_vma:
+        * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
+        * above cannot corrupt).
+        */
+       if (!page_mapped(page)) {
+               put_anon_vma(anon_vma);
+               anon_vma = NULL;
+       }
+out:
+       rcu_read_unlock();
+
+       return anon_vma;
+}
+
+/*
+ * Similar to page_get_anon_vma() except it locks the anon_vma.
+ *
+ * Its a little more complex as it tries to keep the fast path to a single
+ * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
+ * reference like with page_get_anon_vma() and then block on the mutex.
  */
 struct anon_vma *page_lock_anon_vma(struct page *page)
 {
-       struct anon_vma *anon_vma;
+       struct anon_vma *anon_vma = NULL;
+       struct anon_vma *root_anon_vma;
        unsigned long anon_mapping;
 
        rcu_read_lock();
-       anon_mapping = (unsigned long) page->mapping;
-       if (!(anon_mapping & PAGE_MAPPING_ANON))
+       anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
+       if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
                goto out;
        if (!page_mapped(page))
                goto out;
 
        anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
-       spin_lock(&anon_vma->lock);
+       root_anon_vma = ACCESS_ONCE(anon_vma->root);
+       if (mutex_trylock(&root_anon_vma->mutex)) {
+               /*
+                * If the page is still mapped, then this anon_vma is still
+                * its anon_vma, and holding the mutex ensures that it will
+                * not go away, see anon_vma_free().
+                */
+               if (!page_mapped(page)) {
+                       mutex_unlock(&root_anon_vma->mutex);
+                       anon_vma = NULL;
+               }
+               goto out;
+       }
+
+       /* trylock failed, we got to sleep */
+       if (!atomic_inc_not_zero(&anon_vma->refcount)) {
+               anon_vma = NULL;
+               goto out;
+       }
+
+       if (!page_mapped(page)) {
+               put_anon_vma(anon_vma);
+               anon_vma = NULL;
+               goto out;
+       }
+
+       /* we pinned the anon_vma, its safe to sleep */
+       rcu_read_unlock();
+       anon_vma_lock(anon_vma);
+
+       if (atomic_dec_and_test(&anon_vma->refcount)) {
+               /*
+                * Oops, we held the last refcount, release the lock
+                * and bail -- can't simply use put_anon_vma() because
+                * we'll deadlock on the anon_vma_lock() recursion.
+                */
+               anon_vma_unlock(anon_vma);
+               __put_anon_vma(anon_vma);
+               anon_vma = NULL;
+       }
+
        return anon_vma;
+
 out:
        rcu_read_unlock();
-       return NULL;
+       return anon_vma;
 }
 
 void page_unlock_anon_vma(struct anon_vma *anon_vma)
 {
-       spin_unlock(&anon_vma->lock);
-       rcu_read_unlock();
+       anon_vma_unlock(anon_vma);
 }
 
 /*
@@ -227,12 +564,14 @@ void page_unlock_anon_vma(struct anon_vma *anon_vma)
  * Returns virtual address or -EFAULT if page's index/offset is not
  * within the range mapped the @vma.
  */
-static inline unsigned long
+inline unsigned long
 vma_address(struct page *page, struct vm_area_struct *vma)
 {
        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
        unsigned long address;
 
+       if (unlikely(is_vm_hugetlb_page(vma)))
+               pgoff = page->index << huge_page_order(page_hstate(page));
        address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
        if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
                /* page should be within @vma mapping range */
@@ -243,13 +582,18 @@ vma_address(struct page *page, struct vm_area_struct *vma)
 
 /*
  * At what user virtual address is page expected in vma?
- * checking that the page matches the vma.
+ * Caller should check the page is actually part of the vma.
  */
 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
 {
        if (PageAnon(page)) {
-               if ((void *)vma->anon_vma !=
-                   (void *)page->mapping - PAGE_MAPPING_ANON)
+               struct anon_vma *page__anon_vma = page_anon_vma(page);
+               /*
+                * Note: swapoff's unuse_vma() is more efficient with this
+                * check, and needs it to match anon_vma when KSM is active.
+                */
+               if (!vma->anon_vma || !page__anon_vma ||
+                   vma->anon_vma->root != page__anon_vma->root)
                        return -EFAULT;
        } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
                if (!vma->vm_file ||
@@ -269,7 +613,7 @@ unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
  *
  * On success returns with pte mapped and locked.
  */
-pte_t *page_check_address(struct page *page, struct mm_struct *mm,
+pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
                          unsigned long address, spinlock_t **ptlp, int sync)
 {
        pgd_t *pgd;
@@ -278,6 +622,12 @@ pte_t *page_check_address(struct page *page, struct mm_struct *mm,
        pte_t *pte;
        spinlock_t *ptl;
 
+       if (unlikely(PageHuge(page))) {
+               pte = huge_pte_offset(mm, address);
+               ptl = &mm->page_table_lock;
+               goto check;
+       }
+
        pgd = pgd_offset(mm, address);
        if (!pgd_present(*pgd))
                return NULL;
@@ -289,6 +639,8 @@ pte_t *page_check_address(struct page *page, struct mm_struct *mm,
        pmd = pmd_offset(pud, address);
        if (!pmd_present(*pmd))
                return NULL;
+       if (pmd_trans_huge(*pmd))
+               return NULL;
 
        pte = pte_offset_map(pmd, address);
        /* Make a quick check before getting the lock */
@@ -298,6 +650,7 @@ pte_t *page_check_address(struct page *page, struct mm_struct *mm,
        }
 
        ptl = pte_lockptr(mm, pmd);
+check:
        spin_lock(ptl);
        if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
                *ptlp = ptl;
@@ -337,46 +690,70 @@ int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
  * Subfunctions of page_referenced: page_referenced_one called
  * repeatedly from either page_referenced_anon or page_referenced_file.
  */
-static int page_referenced_one(struct page *page,
-                              struct vm_area_struct *vma,
-                              unsigned int *mapcount,
-                              unsigned long *vm_flags)
+int page_referenced_one(struct page *page, struct vm_area_struct *vma,
+                       unsigned long address, unsigned int *mapcount,
+                       unsigned long *vm_flags)
 {
        struct mm_struct *mm = vma->vm_mm;
-       unsigned long address;
-       pte_t *pte;
-       spinlock_t *ptl;
        int referenced = 0;
 
-       address = vma_address(page, vma);
-       if (address == -EFAULT)
-               goto out;
+       if (unlikely(PageTransHuge(page))) {
+               pmd_t *pmd;
 
-       pte = page_check_address(page, mm, address, &ptl, 0);
-       if (!pte)
-               goto out;
+               spin_lock(&mm->page_table_lock);
+               /*
+                * rmap might return false positives; we must filter
+                * these out using page_check_address_pmd().
+                */
+               pmd = page_check_address_pmd(page, mm, address,
+                                            PAGE_CHECK_ADDRESS_PMD_FLAG);
+               if (!pmd) {
+                       spin_unlock(&mm->page_table_lock);
+                       goto out;
+               }
 
-       /*
-        * Don't want to elevate referenced for mlocked page that gets this far,
-        * in order that it progresses to try_to_unmap and is moved to the
-        * unevictable list.
-        */
-       if (vma->vm_flags & VM_LOCKED) {
-               *mapcount = 1;  /* break early from loop */
-               *vm_flags |= VM_LOCKED;
-               goto out_unmap;
-       }
+               if (vma->vm_flags & VM_LOCKED) {
+                       spin_unlock(&mm->page_table_lock);
+                       *mapcount = 0;  /* break early from loop */
+                       *vm_flags |= VM_LOCKED;
+                       goto out;
+               }
+
+               /* go ahead even if the pmd is pmd_trans_splitting() */
+               if (pmdp_clear_flush_young_notify(vma, address, pmd))
+                       referenced++;
+               spin_unlock(&mm->page_table_lock);
+       } else {
+               pte_t *pte;
+               spinlock_t *ptl;
 
-       if (ptep_clear_flush_young_notify(vma, address, pte)) {
                /*
-                * Don't treat a reference through a sequentially read
-                * mapping as such.  If the page has been used in
-                * another mapping, we will catch it; if this other
-                * mapping is already gone, the unmap path will have
-                * set PG_referenced or activated the page.
+                * rmap might return false positives; we must filter
+                * these out using page_check_address().
                 */
-               if (likely(!VM_SequentialReadHint(vma)))
-                       referenced++;
+               pte = page_check_address(page, mm, address, &ptl, 0);
+               if (!pte)
+                       goto out;
+
+               if (vma->vm_flags & VM_LOCKED) {
+                       pte_unmap_unlock(pte, ptl);
+                       *mapcount = 0;  /* break early from loop */
+                       *vm_flags |= VM_LOCKED;
+                       goto out;
+               }
+
+               if (ptep_clear_flush_young_notify(vma, address, pte)) {
+                       /*
+                        * Don't treat a reference through a sequentially read
+                        * mapping as such.  If the page has been used in
+                        * another mapping, we will catch it; if this other
+                        * mapping is already gone, the unmap path will have
+                        * set PG_referenced or activated the page.
+                        */
+                       if (likely(!VM_SequentialReadHint(vma)))
+                               referenced++;
+               }
+               pte_unmap_unlock(pte, ptl);
        }
 
        /* Pretend the page is referenced if the task has the
@@ -385,22 +762,21 @@ static int page_referenced_one(struct page *page,
                        rwsem_is_locked(&mm->mmap_sem))
                referenced++;
 
-out_unmap:
        (*mapcount)--;
-       pte_unmap_unlock(pte, ptl);
-out:
+
        if (referenced)
                *vm_flags |= vma->vm_flags;
+out:
        return referenced;
 }
 
 static int page_referenced_anon(struct page *page,
-                               struct mem_cgroup *mem_cont,
+                               struct mem_cgroup *memcg,
                                unsigned long *vm_flags)
 {
        unsigned int mapcount;
        struct anon_vma *anon_vma;
-       struct vm_area_struct *vma;
+       struct anon_vma_chain *avc;
        int referenced = 0;
 
        anon_vma = page_lock_anon_vma(page);
@@ -408,15 +784,19 @@ static int page_referenced_anon(struct page *page,
                return referenced;
 
        mapcount = page_mapcount(page);
-       list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
+       list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
+               struct vm_area_struct *vma = avc->vma;
+               unsigned long address = vma_address(page, vma);
+               if (address == -EFAULT)
+                       continue;
                /*
                 * If we are reclaiming on behalf of a cgroup, skip
                 * counting on behalf of references from different
                 * cgroups
                 */
-               if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
+               if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
                        continue;
-               referenced += page_referenced_one(page, vma,
+               referenced += page_referenced_one(page, vma, address,
                                                  &mapcount, vm_flags);
                if (!mapcount)
                        break;
@@ -429,7 +809,7 @@ static int page_referenced_anon(struct page *page,
 /**
  * page_referenced_file - referenced check for object-based rmap
  * @page: the page we're checking references on.
- * @mem_cont: target memory controller
+ * @memcg: target memory control group
  * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
  *
  * For an object-based mapped page, find all the places it is mapped and
@@ -440,7 +820,7 @@ static int page_referenced_anon(struct page *page,
  * This function is only called from page_referenced for object-based pages.
  */
 static int page_referenced_file(struct page *page,
-                               struct mem_cgroup *mem_cont,
+                               struct mem_cgroup *memcg,
                                unsigned long *vm_flags)
 {
        unsigned int mapcount;
@@ -461,33 +841,36 @@ static int page_referenced_file(struct page *page,
         * The page lock not only makes sure that page->mapping cannot
         * suddenly be NULLified by truncation, it makes sure that the
         * structure at mapping cannot be freed and reused yet,
-        * so we can safely take mapping->i_mmap_lock.
+        * so we can safely take mapping->i_mmap_mutex.
         */
        BUG_ON(!PageLocked(page));
 
-       spin_lock(&mapping->i_mmap_lock);
+       mutex_lock(&mapping->i_mmap_mutex);
 
        /*
-        * i_mmap_lock does not stabilize mapcount at all, but mapcount
+        * i_mmap_mutex does not stabilize mapcount at all, but mapcount
         * is more likely to be accurate if we note it after spinning.
         */
        mapcount = page_mapcount(page);
 
        vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
+               unsigned long address = vma_address(page, vma);
+               if (address == -EFAULT)
+                       continue;
                /*
                 * If we are reclaiming on behalf of a cgroup, skip
                 * counting on behalf of references from different
                 * cgroups
                 */
-               if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
+               if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
                        continue;
-               referenced += page_referenced_one(page, vma,
+               referenced += page_referenced_one(page, vma, address,
                                                  &mapcount, vm_flags);
                if (!mapcount)
                        break;
        }
 
-       spin_unlock(&mapping->i_mmap_lock);
+       mutex_unlock(&mapping->i_mmap_mutex);
        return referenced;
 }
 
@@ -495,7 +878,7 @@ static int page_referenced_file(struct page *page,
  * page_referenced - test if the page was referenced
  * @page: the page to test
  * @is_locked: caller holds lock on the page
- * @mem_cont: target memory controller
+ * @memcg: target memory cgroup
  * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
  *
  * Quick test_and_clear_referenced for all mappings to a page,
@@ -503,50 +886,48 @@ static int page_referenced_file(struct page *page,
  */
 int page_referenced(struct page *page,
                    int is_locked,
-                   struct mem_cgroup *mem_cont,
+                   struct mem_cgroup *memcg,
                    unsigned long *vm_flags)
 {
        int referenced = 0;
-
-       if (TestClearPageReferenced(page))
-               referenced++;
+       int we_locked = 0;
 
        *vm_flags = 0;
-       if (page_mapped(page) && page->mapping) {
-               if (PageAnon(page))
-                       referenced += page_referenced_anon(page, mem_cont,
+       if (page_mapped(page) && page_rmapping(page)) {
+               if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
+                       we_locked = trylock_page(page);
+                       if (!we_locked) {
+                               referenced++;
+                               goto out;
+                       }
+               }
+               if (unlikely(PageKsm(page)))
+                       referenced += page_referenced_ksm(page, memcg,
                                                                vm_flags);
-               else if (is_locked)
-                       referenced += page_referenced_file(page, mem_cont,
+               else if (PageAnon(page))
+                       referenced += page_referenced_anon(page, memcg,
                                                                vm_flags);
-               else if (!trylock_page(page))
-                       referenced++;
-               else {
-                       if (page->mapping)
-                               referenced += page_referenced_file(page,
-                                                       mem_cont, vm_flags);
+               else if (page->mapping)
+                       referenced += page_referenced_file(page, memcg,
+                                                               vm_flags);
+               if (we_locked)
                        unlock_page(page);
-               }
-       }
-
-       if (page_test_and_clear_young(page))
-               referenced++;
 
+               if (page_test_and_clear_young(page_to_pfn(page)))
+                       referenced++;
+       }
+out:
        return referenced;
 }
 
-static int page_mkclean_one(struct page *page, struct vm_area_struct *vma)
+static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
+                           unsigned long address)
 {
        struct mm_struct *mm = vma->vm_mm;
-       unsigned long address;
        pte_t *pte;
        spinlock_t *ptl;
        int ret = 0;
 
-       address = vma_address(page, vma);
-       if (address == -EFAULT)
-               goto out;
-
        pte = page_check_address(page, mm, address, &ptl, 1);
        if (!pte)
                goto out;
@@ -576,12 +957,16 @@ static int page_mkclean_file(struct address_space *mapping, struct page *page)
 
        BUG_ON(PageAnon(page));
 
-       spin_lock(&mapping->i_mmap_lock);
+       mutex_lock(&mapping->i_mmap_mutex);
        vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
-               if (vma->vm_flags & VM_SHARED)
-                       ret += page_mkclean_one(page, vma);
+               if (vma->vm_flags & VM_SHARED) {
+                       unsigned long address = vma_address(page, vma);
+                       if (address == -EFAULT)
+                               continue;
+                       ret += page_mkclean_one(page, vma, address);
+               }
        }
-       spin_unlock(&mapping->i_mmap_lock);
+       mutex_unlock(&mapping->i_mmap_mutex);
        return ret;
 }
 
@@ -593,13 +978,8 @@ int page_mkclean(struct page *page)
 
        if (page_mapped(page)) {
                struct address_space *mapping = page_mapping(page);
-               if (mapping) {
+               if (mapping)
                        ret = page_mkclean_file(mapping, page);
-                       if (page_test_dirty(page)) {
-                               page_clear_dirty(page);
-                               ret = 1;
-                       }
-               }
        }
 
        return ret;
@@ -607,27 +987,57 @@ int page_mkclean(struct page *page)
 EXPORT_SYMBOL_GPL(page_mkclean);
 
 /**
- * __page_set_anon_rmap - setup new anonymous rmap
- * @page:      the page to add the mapping to
- * @vma:       the vm area in which the mapping is added
+ * page_move_anon_rmap - move a page to our anon_vma
+ * @page:      the page to move to our anon_vma
+ * @vma:       the vma the page belongs to
  * @address:   the user virtual address mapped
+ *
+ * When a page belongs exclusively to one process after a COW event,
+ * that page can be moved into the anon_vma that belongs to just that
+ * process, so the rmap code will not search the parent or sibling
+ * processes.
  */
-static void __page_set_anon_rmap(struct page *page,
+void page_move_anon_rmap(struct page *page,
        struct vm_area_struct *vma, unsigned long address)
 {
        struct anon_vma *anon_vma = vma->anon_vma;
 
-       BUG_ON(!anon_vma);
+       VM_BUG_ON(!PageLocked(page));
+       VM_BUG_ON(!anon_vma);
+       VM_BUG_ON(page->index != linear_page_index(vma, address));
+
        anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
        page->mapping = (struct address_space *) anon_vma;
+}
 
-       page->index = linear_page_index(vma, address);
+/**
+ * __page_set_anon_rmap - set up new anonymous rmap
+ * @page:      Page to add to rmap     
+ * @vma:       VM area to add page to.
+ * @address:   User virtual address of the mapping     
+ * @exclusive: the page is exclusively owned by the current process
+ */
+static void __page_set_anon_rmap(struct page *page,
+       struct vm_area_struct *vma, unsigned long address, int exclusive)
+{
+       struct anon_vma *anon_vma = vma->anon_vma;
+
+       BUG_ON(!anon_vma);
+
+       if (PageAnon(page))
+               return;
 
        /*
-        * nr_mapped state can be updated without turning off
-        * interrupts because it is not modified via interrupt.
+        * If the page isn't exclusively mapped into this vma,
+        * we must use the _oldest_ possible anon_vma for the
+        * page mapping!
         */
-       __inc_zone_page_state(page, NR_ANON_PAGES);
+       if (!exclusive)
+               anon_vma = anon_vma->root;
+
+       anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
+       page->mapping = (struct address_space *) anon_vma;
+       page->index = linear_page_index(vma, address);
 }
 
 /**
@@ -652,9 +1062,7 @@ static void __page_check_anon_rmap(struct page *page,
         * are initially only visible via the pagetables, and the pte is locked
         * over the call to page_add_new_anon_rmap.
         */
-       struct anon_vma *anon_vma = vma->anon_vma;
-       anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
-       BUG_ON(page->mapping != (struct address_space *)anon_vma);
+       BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
        BUG_ON(page->index != linear_page_index(vma, address));
 #endif
 }
@@ -665,15 +1073,40 @@ static void __page_check_anon_rmap(struct page *page,
  * @vma:       the vm area in which the mapping is added
  * @address:   the user virtual address mapped
  *
- * The caller needs to hold the pte lock and the page must be locked.
+ * The caller needs to hold the pte lock, and the page must be locked in
+ * the anon_vma case: to serialize mapping,index checking after setting,
+ * and to ensure that PageAnon is not being upgraded racily to PageKsm
+ * (but PageKsm is never downgraded to PageAnon).
  */
 void page_add_anon_rmap(struct page *page,
        struct vm_area_struct *vma, unsigned long address)
 {
+       do_page_add_anon_rmap(page, vma, address, 0);
+}
+
+/*
+ * Special version of the above for do_swap_page, which often runs
+ * into pages that are exclusively owned by the current process.
+ * Everybody else should continue to use page_add_anon_rmap above.
+ */
+void do_page_add_anon_rmap(struct page *page,
+       struct vm_area_struct *vma, unsigned long address, int exclusive)
+{
+       int first = atomic_inc_and_test(&page->_mapcount);
+       if (first) {
+               if (!PageTransHuge(page))
+                       __inc_zone_page_state(page, NR_ANON_PAGES);
+               else
+                       __inc_zone_page_state(page,
+                                             NR_ANON_TRANSPARENT_HUGEPAGES);
+       }
+       if (unlikely(PageKsm(page)))
+               return;
+
        VM_BUG_ON(!PageLocked(page));
-       VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
-       if (atomic_inc_and_test(&page->_mapcount))
-               __page_set_anon_rmap(page, vma, address);
+       /* address might be in next vma when migration races vma_adjust */
+       if (first)
+               __page_set_anon_rmap(page, vma, address, exclusive);
        else
                __page_check_anon_rmap(page, vma, address);
 }
@@ -694,7 +1127,11 @@ void page_add_new_anon_rmap(struct page *page,
        VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
        SetPageSwapBacked(page);
        atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
-       __page_set_anon_rmap(page, vma, address);
+       if (!PageTransHuge(page))
+               __inc_zone_page_state(page, NR_ANON_PAGES);
+       else
+               __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
+       __page_set_anon_rmap(page, vma, address, 1);
        if (page_evictable(page, vma))
                lru_cache_add_lru(page, LRU_ACTIVE_ANON);
        else
@@ -709,10 +1146,15 @@ void page_add_new_anon_rmap(struct page *page,
  */
 void page_add_file_rmap(struct page *page)
 {
+       bool locked;
+       unsigned long flags;
+
+       mem_cgroup_begin_update_page_stat(page, &locked, &flags);
        if (atomic_inc_and_test(&page->_mapcount)) {
                __inc_zone_page_state(page, NR_FILE_MAPPED);
-               mem_cgroup_update_mapped_file_stat(page, 1);
+               mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED);
        }
+       mem_cgroup_end_update_page_stat(page, &locked, &flags);
 }
 
 /**
@@ -723,9 +1165,22 @@ void page_add_file_rmap(struct page *page)
  */
 void page_remove_rmap(struct page *page)
 {
+       struct address_space *mapping = page_mapping(page);
+       bool anon = PageAnon(page);
+       bool locked;
+       unsigned long flags;
+
+       /*
+        * The anon case has no mem_cgroup page_stat to update; but may
+        * uncharge_page() below, where the lock ordering can deadlock if
+        * we hold the lock against page_stat move: so avoid it on anon.
+        */
+       if (!anon)
+               mem_cgroup_begin_update_page_stat(page, &locked, &flags);
+
        /* page still mapped by someone else? */
        if (!atomic_add_negative(-1, &page->_mapcount))
-               return;
+               goto out;
 
        /*
         * Now that the last pte has gone, s390 must transfer dirty
@@ -733,18 +1188,38 @@ void page_remove_rmap(struct page *page)
         * this if the page is anon, so about to be freed; but perhaps
         * not if it's in swapcache - there might be another pte slot
         * containing the swap entry, but page not yet written to swap.
+        *
+        * And we can skip it on file pages, so long as the filesystem
+        * participates in dirty tracking; but need to catch shm and tmpfs
+        * and ramfs pages which have been modified since creation by read
+        * fault.
+        *
+        * Note that mapping must be decided above, before decrementing
+        * mapcount (which luckily provides a barrier): once page is unmapped,
+        * it could be truncated and page->mapping reset to NULL at any moment.
+        * Note also that we are relying on page_mapping(page) to set mapping
+        * to &swapper_space when PageSwapCache(page).
         */
-       if ((!PageAnon(page) || PageSwapCache(page)) && page_test_dirty(page)) {
-               page_clear_dirty(page);
+       if (mapping && !mapping_cap_account_dirty(mapping) &&
+           page_test_and_clear_dirty(page_to_pfn(page), 1))
                set_page_dirty(page);
-       }
-       if (PageAnon(page)) {
+       /*
+        * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
+        * and not charged by memcg for now.
+        */
+       if (unlikely(PageHuge(page)))
+               goto out;
+       if (anon) {
                mem_cgroup_uncharge_page(page);
-               __dec_zone_page_state(page, NR_ANON_PAGES);
+               if (!PageTransHuge(page))
+                       __dec_zone_page_state(page, NR_ANON_PAGES);
+               else
+                       __dec_zone_page_state(page,
+                                             NR_ANON_TRANSPARENT_HUGEPAGES);
        } else {
                __dec_zone_page_state(page, NR_FILE_MAPPED);
+               mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED);
        }
-       mem_cgroup_update_mapped_file_stat(page, -1);
        /*
         * It would be tidy to reset the PageAnon mapping here,
         * but that might overwrite a racing page_add_anon_rmap
@@ -754,26 +1229,24 @@ void page_remove_rmap(struct page *page)
         * Leaving it set also helps swapoff to reinstate ptes
         * faster for those pages still in swapcache.
         */
+out:
+       if (!anon)
+               mem_cgroup_end_update_page_stat(page, &locked, &flags);
 }
 
 /*
  * Subfunctions of try_to_unmap: try_to_unmap_one called
- * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
+ * repeatedly from try_to_unmap_ksm, try_to_unmap_anon or try_to_unmap_file.
  */
-static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
-                               enum ttu_flags flags)
+int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
+                    unsigned long address, enum ttu_flags flags)
 {
        struct mm_struct *mm = vma->vm_mm;
-       unsigned long address;
        pte_t *pte;
        pte_t pteval;
        spinlock_t *ptl;
        int ret = SWAP_AGAIN;
 
-       address = vma_address(page, vma);
-       if (address == -EFAULT)
-               goto out;
-
        pte = page_check_address(page, mm, address, &ptl, 0);
        if (!pte)
                goto out;
@@ -784,10 +1257,11 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
         * skipped over this mm) then we should reactivate it.
         */
        if (!(flags & TTU_IGNORE_MLOCK)) {
-               if (vma->vm_flags & VM_LOCKED) {
-                       ret = SWAP_MLOCK;
+               if (vma->vm_flags & VM_LOCKED)
+                       goto out_mlock;
+
+               if (TTU_ACTION(flags) == TTU_MUNLOCK)
                        goto out_unmap;
-               }
        }
        if (!(flags & TTU_IGNORE_ACCESS)) {
                if (ptep_clear_flush_young_notify(vma, address, pte)) {
@@ -809,9 +1283,9 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
 
        if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
                if (PageAnon(page))
-                       dec_mm_counter(mm, anon_rss);
+                       dec_mm_counter(mm, MM_ANONPAGES);
                else
-                       dec_mm_counter(mm, file_rss);
+                       dec_mm_counter(mm, MM_FILEPAGES);
                set_pte_at(mm, address, pte,
                                swp_entry_to_pte(make_hwpoison_entry(page)));
        } else if (PageAnon(page)) {
@@ -822,15 +1296,20 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
                         * Store the swap location in the pte.
                         * See handle_pte_fault() ...
                         */
-                       swap_duplicate(entry);
+                       if (swap_duplicate(entry) < 0) {
+                               set_pte_at(mm, address, pte, pteval);
+                               ret = SWAP_FAIL;
+                               goto out_unmap;
+                       }
                        if (list_empty(&mm->mmlist)) {
                                spin_lock(&mmlist_lock);
                                if (list_empty(&mm->mmlist))
                                        list_add(&mm->mmlist, &init_mm.mmlist);
                                spin_unlock(&mmlist_lock);
                        }
-                       dec_mm_counter(mm, anon_rss);
-               } else if (PAGE_MIGRATION) {
+                       dec_mm_counter(mm, MM_ANONPAGES);
+                       inc_mm_counter(mm, MM_SWAPENTS);
+               } else if (IS_ENABLED(CONFIG_MIGRATION)) {
                        /*
                         * Store the pfn of the page in a special migration
                         * pte. do_swap_page() will wait until the migration
@@ -841,14 +1320,14 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
                }
                set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
                BUG_ON(pte_file(*pte));
-       } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
+       } else if (IS_ENABLED(CONFIG_MIGRATION) &&
+                  (TTU_ACTION(flags) == TTU_MIGRATION)) {
                /* Establish migration entry for a file page */
                swp_entry_t entry;
                entry = make_migration_entry(page, pte_write(pteval));
                set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
        } else
-               dec_mm_counter(mm, file_rss);
-
+               dec_mm_counter(mm, MM_FILEPAGES);
 
        page_remove_rmap(page);
        page_cache_release(page);
@@ -857,6 +1336,27 @@ out_unmap:
        pte_unmap_unlock(pte, ptl);
 out:
        return ret;
+
+out_mlock:
+       pte_unmap_unlock(pte, ptl);
+
+
+       /*
+        * We need mmap_sem locking, Otherwise VM_LOCKED check makes
+        * unstable result and race. Plus, We can't wait here because
+        * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
+        * if trylock failed, the page remain in evictable lru and later
+        * vmscan could retry to move the page to unevictable lru if the
+        * page is actually mlocked.
+        */
+       if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
+               if (vma->vm_flags & VM_LOCKED) {
+                       mlock_vma_page(page);
+                       ret = SWAP_MLOCK;
+               }
+               up_read(&vma->vm_mm->mmap_sem);
+       }
+       return ret;
 }
 
 /*
@@ -922,11 +1422,10 @@ static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
                return ret;
 
        /*
-        * MLOCK_PAGES => feature is configured.
-        * if we can acquire the mmap_sem for read, and vma is VM_LOCKED,
+        * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
         * keep the sem while scanning the cluster for mlocking pages.
         */
-       if (MLOCK_PAGES && down_read_trylock(&vma->vm_mm->mmap_sem)) {
+       if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
                locked_vma = (vma->vm_flags & VM_LOCKED);
                if (!locked_vma)
                        up_read(&vma->vm_mm->mmap_sem); /* don't need it */
@@ -967,7 +1466,7 @@ static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
 
                page_remove_rmap(page);
                page_cache_release(page);
-               dec_mm_counter(mm, file_rss);
+               dec_mm_counter(mm, MM_FILEPAGES);
                (*mapcount)--;
        }
        pte_unmap_unlock(pte - 1, ptl);
@@ -976,29 +1475,25 @@ static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
        return ret;
 }
 
-/*
- * common handling for pages mapped in VM_LOCKED vmas
- */
-static int try_to_mlock_page(struct page *page, struct vm_area_struct *vma)
+bool is_vma_temporary_stack(struct vm_area_struct *vma)
 {
-       int mlocked = 0;
+       int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
 
-       if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
-               if (vma->vm_flags & VM_LOCKED) {
-                       mlock_vma_page(page);
-                       mlocked++;      /* really mlocked the page */
-               }
-               up_read(&vma->vm_mm->mmap_sem);
-       }
-       return mlocked;
+       if (!maybe_stack)
+               return false;
+
+       if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
+                                               VM_STACK_INCOMPLETE_SETUP)
+               return true;
+
+       return false;
 }
 
 /**
  * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
  * rmap method
  * @page: the page to unmap/unlock
- * @unlock:  request for unlock rather than unmap [unlikely]
- * @migration:  unmapping for migration - ignored if @unlock
+ * @flags: action and flags
  *
  * Find all the mappings of a page using the mapping pointer and the vma chains
  * contained in the anon_vma struct it points to.
@@ -1013,43 +1508,38 @@ static int try_to_mlock_page(struct page *page, struct vm_area_struct *vma)
 static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
 {
        struct anon_vma *anon_vma;
-       struct vm_area_struct *vma;
-       unsigned int mlocked = 0;
+       struct anon_vma_chain *avc;
        int ret = SWAP_AGAIN;
-       int unlock = TTU_ACTION(flags) == TTU_MUNLOCK;
-
-       if (MLOCK_PAGES && unlikely(unlock))
-               ret = SWAP_SUCCESS;     /* default for try_to_munlock() */
 
        anon_vma = page_lock_anon_vma(page);
        if (!anon_vma)
                return ret;
 
-       list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
-               if (MLOCK_PAGES && unlikely(unlock)) {
-                       if (!((vma->vm_flags & VM_LOCKED) &&
-                             page_mapped_in_vma(page, vma)))
-                               continue;  /* must visit all unlocked vmas */
-                       ret = SWAP_MLOCK;  /* saw at least one mlocked vma */
-               } else {
-                       ret = try_to_unmap_one(page, vma, flags);
-                       if (ret == SWAP_FAIL || !page_mapped(page))
-                               break;
-               }
-               if (ret == SWAP_MLOCK) {
-                       mlocked = try_to_mlock_page(page, vma);
-                       if (mlocked)
-                               break;  /* stop if actually mlocked page */
-               }
-       }
+       list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
+               struct vm_area_struct *vma = avc->vma;
+               unsigned long address;
 
-       page_unlock_anon_vma(anon_vma);
+               /*
+                * During exec, a temporary VMA is setup and later moved.
+                * The VMA is moved under the anon_vma lock but not the
+                * page tables leading to a race where migration cannot
+                * find the migration ptes. Rather than increasing the
+                * locking requirements of exec(), migration skips
+                * temporary VMAs until after exec() completes.
+                */
+               if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
+                               is_vma_temporary_stack(vma))
+                       continue;
 
-       if (mlocked)
-               ret = SWAP_MLOCK;       /* actually mlocked the page */
-       else if (ret == SWAP_MLOCK)
-               ret = SWAP_AGAIN;       /* saw VM_LOCKED vma */
+               address = vma_address(page, vma);
+               if (address == -EFAULT)
+                       continue;
+               ret = try_to_unmap_one(page, vma, address, flags);
+               if (ret != SWAP_AGAIN || !page_mapped(page))
+                       break;
+       }
 
+       page_unlock_anon_vma(anon_vma);
        return ret;
 }
 
@@ -1079,48 +1569,30 @@ static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
        unsigned long max_nl_cursor = 0;
        unsigned long max_nl_size = 0;
        unsigned int mapcount;
-       unsigned int mlocked = 0;
-       int unlock = TTU_ACTION(flags) == TTU_MUNLOCK;
-
-       if (MLOCK_PAGES && unlikely(unlock))
-               ret = SWAP_SUCCESS;     /* default for try_to_munlock() */
 
-       spin_lock(&mapping->i_mmap_lock);
+       mutex_lock(&mapping->i_mmap_mutex);
        vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
-               if (MLOCK_PAGES && unlikely(unlock)) {
-                       if (!((vma->vm_flags & VM_LOCKED) &&
-                                               page_mapped_in_vma(page, vma)))
-                               continue;       /* must visit all vmas */
-                       ret = SWAP_MLOCK;
-               } else {
-                       ret = try_to_unmap_one(page, vma, flags);
-                       if (ret == SWAP_FAIL || !page_mapped(page))
-                               goto out;
-               }
-               if (ret == SWAP_MLOCK) {
-                       mlocked = try_to_mlock_page(page, vma);
-                       if (mlocked)
-                               break;  /* stop if actually mlocked page */
-               }
+               unsigned long address = vma_address(page, vma);
+               if (address == -EFAULT)
+                       continue;
+               ret = try_to_unmap_one(page, vma, address, flags);
+               if (ret != SWAP_AGAIN || !page_mapped(page))
+                       goto out;
        }
 
-       if (mlocked)
+       if (list_empty(&mapping->i_mmap_nonlinear))
                goto out;
 
-       if (list_empty(&mapping->i_mmap_nonlinear))
+       /*
+        * We don't bother to try to find the munlocked page in nonlinears.
+        * It's costly. Instead, later, page reclaim logic may call
+        * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
+        */
+       if (TTU_ACTION(flags) == TTU_MUNLOCK)
                goto out;
 
        list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
                                                shared.vm_set.list) {
-               if (MLOCK_PAGES && unlikely(unlock)) {
-                       if (!(vma->vm_flags & VM_LOCKED))
-                               continue;       /* must visit all vmas */
-                       ret = SWAP_MLOCK;       /* leave mlocked == 0 */
-                       goto out;               /* no need to look further */
-               }
-               if (!MLOCK_PAGES && !(flags & TTU_IGNORE_MLOCK) &&
-                       (vma->vm_flags & VM_LOCKED))
-                       continue;
                cursor = (unsigned long) vma->vm_private_data;
                if (cursor > max_nl_cursor)
                        max_nl_cursor = cursor;
@@ -1144,7 +1616,7 @@ static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
        mapcount = page_mapcount(page);
        if (!mapcount)
                goto out;
-       cond_resched_lock(&mapping->i_mmap_lock);
+       cond_resched();
 
        max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
        if (max_nl_cursor == 0)
@@ -1153,16 +1625,12 @@ static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
        do {
                list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
                                                shared.vm_set.list) {
-                       if (!MLOCK_PAGES && !(flags & TTU_IGNORE_MLOCK) &&
-                           (vma->vm_flags & VM_LOCKED))
-                               continue;
                        cursor = (unsigned long) vma->vm_private_data;
                        while ( cursor < max_nl_cursor &&
                                cursor < vma->vm_end - vma->vm_start) {
-                               ret = try_to_unmap_cluster(cursor, &mapcount,
-                                                               vma, page);
-                               if (ret == SWAP_MLOCK)
-                                       mlocked = 2;    /* to return below */
+                               if (try_to_unmap_cluster(cursor, &mapcount,
+                                               vma, page) == SWAP_MLOCK)
+                                       ret = SWAP_MLOCK;
                                cursor += CLUSTER_SIZE;
                                vma->vm_private_data = (void *) cursor;
                                if ((int)mapcount <= 0)
@@ -1170,7 +1638,7 @@ static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
                        }
                        vma->vm_private_data = (void *) max_nl_cursor;
                }
-               cond_resched_lock(&mapping->i_mmap_lock);
+               cond_resched();
                max_nl_cursor += CLUSTER_SIZE;
        } while (max_nl_cursor <= max_nl_size);
 
@@ -1182,11 +1650,7 @@ static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
        list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
                vma->vm_private_data = NULL;
 out:
-       spin_unlock(&mapping->i_mmap_lock);
-       if (mlocked)
-               ret = SWAP_MLOCK;       /* actually mlocked the page */
-       else if (ret == SWAP_MLOCK)
-               ret = SWAP_AGAIN;       /* saw VM_LOCKED vma */
+       mutex_unlock(&mapping->i_mmap_mutex);
        return ret;
 }
 
@@ -1209,8 +1673,11 @@ int try_to_unmap(struct page *page, enum ttu_flags flags)
        int ret;
 
        BUG_ON(!PageLocked(page));
+       VM_BUG_ON(!PageHuge(page) && PageTransHuge(page));
 
-       if (PageAnon(page))
+       if (unlikely(PageKsm(page)))
+               ret = try_to_unmap_ksm(page, flags);
+       else if (PageAnon(page))
                ret = try_to_unmap_anon(page, flags);
        else
                ret = try_to_unmap_file(page, flags);
@@ -1229,17 +1696,153 @@ int try_to_unmap(struct page *page, enum ttu_flags flags)
  *
  * Return values are:
  *
- * SWAP_SUCCESS        - no vma's holding page mlocked.
+ * SWAP_AGAIN  - no vma is holding page mlocked, or,
  * SWAP_AGAIN  - page mapped in mlocked vma -- couldn't acquire mmap sem
+ * SWAP_FAIL   - page cannot be located at present
  * SWAP_MLOCK  - page is now mlocked.
  */
 int try_to_munlock(struct page *page)
 {
        VM_BUG_ON(!PageLocked(page) || PageLRU(page));
 
-       if (PageAnon(page))
+       if (unlikely(PageKsm(page)))
+               return try_to_unmap_ksm(page, TTU_MUNLOCK);
+       else if (PageAnon(page))
                return try_to_unmap_anon(page, TTU_MUNLOCK);
        else
                return try_to_unmap_file(page, TTU_MUNLOCK);
 }
 
+void __put_anon_vma(struct anon_vma *anon_vma)
+{
+       struct anon_vma *root = anon_vma->root;
+
+       if (root != anon_vma && atomic_dec_and_test(&root->refcount))
+               anon_vma_free(root);
+
+       anon_vma_free(anon_vma);
+}
+
+#ifdef CONFIG_MIGRATION
+/*
+ * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
+ * Called by migrate.c to remove migration ptes, but might be used more later.
+ */
+static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
+               struct vm_area_struct *, unsigned long, void *), void *arg)
+{
+       struct anon_vma *anon_vma;
+       struct anon_vma_chain *avc;
+       int ret = SWAP_AGAIN;
+
+       /*
+        * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
+        * because that depends on page_mapped(); but not all its usages
+        * are holding mmap_sem. Users without mmap_sem are required to
+        * take a reference count to prevent the anon_vma disappearing
+        */
+       anon_vma = page_anon_vma(page);
+       if (!anon_vma)
+               return ret;
+       anon_vma_lock(anon_vma);
+       list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
+               struct vm_area_struct *vma = avc->vma;
+               unsigned long address = vma_address(page, vma);
+               if (address == -EFAULT)
+                       continue;
+               ret = rmap_one(page, vma, address, arg);
+               if (ret != SWAP_AGAIN)
+                       break;
+       }
+       anon_vma_unlock(anon_vma);
+       return ret;
+}
+
+static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
+               struct vm_area_struct *, unsigned long, void *), void *arg)
+{
+       struct address_space *mapping = page->mapping;
+       pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
+       struct vm_area_struct *vma;
+       struct prio_tree_iter iter;
+       int ret = SWAP_AGAIN;
+
+       if (!mapping)
+               return ret;
+       mutex_lock(&mapping->i_mmap_mutex);
+       vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
+               unsigned long address = vma_address(page, vma);
+               if (address == -EFAULT)
+                       continue;
+               ret = rmap_one(page, vma, address, arg);
+               if (ret != SWAP_AGAIN)
+                       break;
+       }
+       /*
+        * No nonlinear handling: being always shared, nonlinear vmas
+        * never contain migration ptes.  Decide what to do about this
+        * limitation to linear when we need rmap_walk() on nonlinear.
+        */
+       mutex_unlock(&mapping->i_mmap_mutex);
+       return ret;
+}
+
+int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
+               struct vm_area_struct *, unsigned long, void *), void *arg)
+{
+       VM_BUG_ON(!PageLocked(page));
+
+       if (unlikely(PageKsm(page)))
+               return rmap_walk_ksm(page, rmap_one, arg);
+       else if (PageAnon(page))
+               return rmap_walk_anon(page, rmap_one, arg);
+       else
+               return rmap_walk_file(page, rmap_one, arg);
+}
+#endif /* CONFIG_MIGRATION */
+
+#ifdef CONFIG_HUGETLB_PAGE
+/*
+ * The following three functions are for anonymous (private mapped) hugepages.
+ * Unlike common anonymous pages, anonymous hugepages have no accounting code
+ * and no lru code, because we handle hugepages differently from common pages.
+ */
+static void __hugepage_set_anon_rmap(struct page *page,
+       struct vm_area_struct *vma, unsigned long address, int exclusive)
+{
+       struct anon_vma *anon_vma = vma->anon_vma;
+
+       BUG_ON(!anon_vma);
+
+       if (PageAnon(page))
+               return;
+       if (!exclusive)
+               anon_vma = anon_vma->root;
+
+       anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
+       page->mapping = (struct address_space *) anon_vma;
+       page->index = linear_page_index(vma, address);
+}
+
+void hugepage_add_anon_rmap(struct page *page,
+                           struct vm_area_struct *vma, unsigned long address)
+{
+       struct anon_vma *anon_vma = vma->anon_vma;
+       int first;
+
+       BUG_ON(!PageLocked(page));
+       BUG_ON(!anon_vma);
+       /* address might be in next vma when migration races vma_adjust */
+       first = atomic_inc_and_test(&page->_mapcount);
+       if (first)
+               __hugepage_set_anon_rmap(page, vma, address, 0);
+}
+
+void hugepage_add_new_anon_rmap(struct page *page,
+                       struct vm_area_struct *vma, unsigned long address)
+{
+       BUG_ON(address < vma->vm_start || address >= vma->vm_end);
+       atomic_set(&page->_mapcount, 0);
+       __hugepage_set_anon_rmap(page, vma, address, 1);
+}
+#endif /* CONFIG_HUGETLB_PAGE */