mm: sync vmalloc address space page tables in alloc_vm_area()
[linux-2.6.git] / mm / huge_memory.c
index a313403..e2d1587 100644 (file)
 #include <linux/mmu_notifier.h>
 #include <linux/rmap.h>
 #include <linux/swap.h>
+#include <linux/mm_inline.h>
+#include <linux/kthread.h>
+#include <linux/khugepaged.h>
+#include <linux/freezer.h>
+#include <linux/mman.h>
 #include <asm/tlb.h>
 #include <asm/pgalloc.h>
 #include "internal.h"
 
+/*
+ * By default transparent hugepage support is enabled for all mappings
+ * and khugepaged scans all mappings. Defrag is only invoked by
+ * khugepaged hugepage allocations and by page faults inside
+ * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
+ * allocations.
+ */
 unsigned long transparent_hugepage_flags __read_mostly =
-       (1<<TRANSPARENT_HUGEPAGE_FLAG);
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
+       (1<<TRANSPARENT_HUGEPAGE_FLAG)|
+#endif
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
+       (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
+#endif
+       (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
+       (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
+
+/* default scan 8*512 pte (or vmas) every 30 second */
+static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
+static unsigned int khugepaged_pages_collapsed;
+static unsigned int khugepaged_full_scans;
+static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
+/* during fragmentation poll the hugepage allocator once every minute */
+static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
+static struct task_struct *khugepaged_thread __read_mostly;
+static DEFINE_MUTEX(khugepaged_mutex);
+static DEFINE_SPINLOCK(khugepaged_mm_lock);
+static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
+/*
+ * default collapse hugepages if there is at least one pte mapped like
+ * it would have happened if the vma was large enough during page
+ * fault.
+ */
+static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
+
+static int khugepaged(void *none);
+static int mm_slots_hash_init(void);
+static int khugepaged_slab_init(void);
+static void khugepaged_slab_free(void);
+
+#define MM_SLOTS_HASH_HEADS 1024
+static struct hlist_head *mm_slots_hash __read_mostly;
+static struct kmem_cache *mm_slot_cache __read_mostly;
+
+/**
+ * struct mm_slot - hash lookup from mm to mm_slot
+ * @hash: hash collision list
+ * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
+ * @mm: the mm that this information is valid for
+ */
+struct mm_slot {
+       struct hlist_node hash;
+       struct list_head mm_node;
+       struct mm_struct *mm;
+};
+
+/**
+ * struct khugepaged_scan - cursor for scanning
+ * @mm_head: the head of the mm list to scan
+ * @mm_slot: the current mm_slot we are scanning
+ * @address: the next address inside that to be scanned
+ *
+ * There is only the one khugepaged_scan instance of this cursor structure.
+ */
+struct khugepaged_scan {
+       struct list_head mm_head;
+       struct mm_slot *mm_slot;
+       unsigned long address;
+} khugepaged_scan = {
+       .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
+};
+
+
+static int set_recommended_min_free_kbytes(void)
+{
+       struct zone *zone;
+       int nr_zones = 0;
+       unsigned long recommended_min;
+       extern int min_free_kbytes;
+
+       if (!test_bit(TRANSPARENT_HUGEPAGE_FLAG,
+                     &transparent_hugepage_flags) &&
+           !test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
+                     &transparent_hugepage_flags))
+               return 0;
+
+       for_each_populated_zone(zone)
+               nr_zones++;
+
+       /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
+       recommended_min = pageblock_nr_pages * nr_zones * 2;
+
+       /*
+        * Make sure that on average at least two pageblocks are almost free
+        * of another type, one for a migratetype to fall back to and a
+        * second to avoid subsequent fallbacks of other types There are 3
+        * MIGRATE_TYPES we care about.
+        */
+       recommended_min += pageblock_nr_pages * nr_zones *
+                          MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
+
+       /* don't ever allow to reserve more than 5% of the lowmem */
+       recommended_min = min(recommended_min,
+                             (unsigned long) nr_free_buffer_pages() / 20);
+       recommended_min <<= (PAGE_SHIFT-10);
+
+       if (recommended_min > min_free_kbytes)
+               min_free_kbytes = recommended_min;
+       setup_per_zone_wmarks();
+       return 0;
+}
+late_initcall(set_recommended_min_free_kbytes);
+
+static int start_khugepaged(void)
+{
+       int err = 0;
+       if (khugepaged_enabled()) {
+               int wakeup;
+               if (unlikely(!mm_slot_cache || !mm_slots_hash)) {
+                       err = -ENOMEM;
+                       goto out;
+               }
+               mutex_lock(&khugepaged_mutex);
+               if (!khugepaged_thread)
+                       khugepaged_thread = kthread_run(khugepaged, NULL,
+                                                       "khugepaged");
+               if (unlikely(IS_ERR(khugepaged_thread))) {
+                       printk(KERN_ERR
+                              "khugepaged: kthread_run(khugepaged) failed\n");
+                       err = PTR_ERR(khugepaged_thread);
+                       khugepaged_thread = NULL;
+               }
+               wakeup = !list_empty(&khugepaged_scan.mm_head);
+               mutex_unlock(&khugepaged_mutex);
+               if (wakeup)
+                       wake_up_interruptible(&khugepaged_wait);
+
+               set_recommended_min_free_kbytes();
+       } else
+               /* wakeup to exit */
+               wake_up_interruptible(&khugepaged_wait);
+out:
+       return err;
+}
 
 #ifdef CONFIG_SYSFS
+
 static ssize_t double_flag_show(struct kobject *kobj,
                                struct kobj_attribute *attr, char *buf,
                                enum transparent_hugepage_flag enabled,
@@ -68,9 +216,26 @@ static ssize_t enabled_store(struct kobject *kobj,
                             struct kobj_attribute *attr,
                             const char *buf, size_t count)
 {
-       return double_flag_store(kobj, attr, buf, count,
-                                TRANSPARENT_HUGEPAGE_FLAG,
-                                TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
+       ssize_t ret;
+
+       ret = double_flag_store(kobj, attr, buf, count,
+                               TRANSPARENT_HUGEPAGE_FLAG,
+                               TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
+
+       if (ret > 0) {
+               int err = start_khugepaged();
+               if (err)
+                       ret = err;
+       }
+
+       if (ret > 0 &&
+           (test_bit(TRANSPARENT_HUGEPAGE_FLAG,
+                     &transparent_hugepage_flags) ||
+            test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
+                     &transparent_hugepage_flags)))
+               set_recommended_min_free_kbytes();
+
+       return ret;
 }
 static struct kobj_attribute enabled_attr =
        __ATTR(enabled, 0644, enabled_show, enabled_store);
@@ -79,24 +244,28 @@ static ssize_t single_flag_show(struct kobject *kobj,
                                struct kobj_attribute *attr, char *buf,
                                enum transparent_hugepage_flag flag)
 {
-       if (test_bit(flag, &transparent_hugepage_flags))
-               return sprintf(buf, "[yes] no\n");
-       else
-               return sprintf(buf, "yes [no]\n");
+       return sprintf(buf, "%d\n",
+                      !!test_bit(flag, &transparent_hugepage_flags));
 }
+
 static ssize_t single_flag_store(struct kobject *kobj,
                                 struct kobj_attribute *attr,
                                 const char *buf, size_t count,
                                 enum transparent_hugepage_flag flag)
 {
-       if (!memcmp("yes", buf,
-                   min(sizeof("yes")-1, count))) {
+       unsigned long value;
+       int ret;
+
+       ret = kstrtoul(buf, 10, &value);
+       if (ret < 0)
+               return ret;
+       if (value > 1)
+               return -EINVAL;
+
+       if (value)
                set_bit(flag, &transparent_hugepage_flags);
-       } else if (!memcmp("no", buf,
-                          min(sizeof("no")-1, count))) {
+       else
                clear_bit(flag, &transparent_hugepage_flags);
-       } else
-               return -EINVAL;
 
        return count;
 }
@@ -153,20 +322,230 @@ static struct attribute *hugepage_attr[] = {
 
 static struct attribute_group hugepage_attr_group = {
        .attrs = hugepage_attr,
-       .name = "transparent_hugepage",
+};
+
+static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
+                                        struct kobj_attribute *attr,
+                                        char *buf)
+{
+       return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
+}
+
+static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
+                                         struct kobj_attribute *attr,
+                                         const char *buf, size_t count)
+{
+       unsigned long msecs;
+       int err;
+
+       err = strict_strtoul(buf, 10, &msecs);
+       if (err || msecs > UINT_MAX)
+               return -EINVAL;
+
+       khugepaged_scan_sleep_millisecs = msecs;
+       wake_up_interruptible(&khugepaged_wait);
+
+       return count;
+}
+static struct kobj_attribute scan_sleep_millisecs_attr =
+       __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
+              scan_sleep_millisecs_store);
+
+static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
+                                         struct kobj_attribute *attr,
+                                         char *buf)
+{
+       return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
+}
+
+static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
+                                          struct kobj_attribute *attr,
+                                          const char *buf, size_t count)
+{
+       unsigned long msecs;
+       int err;
+
+       err = strict_strtoul(buf, 10, &msecs);
+       if (err || msecs > UINT_MAX)
+               return -EINVAL;
+
+       khugepaged_alloc_sleep_millisecs = msecs;
+       wake_up_interruptible(&khugepaged_wait);
+
+       return count;
+}
+static struct kobj_attribute alloc_sleep_millisecs_attr =
+       __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
+              alloc_sleep_millisecs_store);
+
+static ssize_t pages_to_scan_show(struct kobject *kobj,
+                                 struct kobj_attribute *attr,
+                                 char *buf)
+{
+       return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
+}
+static ssize_t pages_to_scan_store(struct kobject *kobj,
+                                  struct kobj_attribute *attr,
+                                  const char *buf, size_t count)
+{
+       int err;
+       unsigned long pages;
+
+       err = strict_strtoul(buf, 10, &pages);
+       if (err || !pages || pages > UINT_MAX)
+               return -EINVAL;
+
+       khugepaged_pages_to_scan = pages;
+
+       return count;
+}
+static struct kobj_attribute pages_to_scan_attr =
+       __ATTR(pages_to_scan, 0644, pages_to_scan_show,
+              pages_to_scan_store);
+
+static ssize_t pages_collapsed_show(struct kobject *kobj,
+                                   struct kobj_attribute *attr,
+                                   char *buf)
+{
+       return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
+}
+static struct kobj_attribute pages_collapsed_attr =
+       __ATTR_RO(pages_collapsed);
+
+static ssize_t full_scans_show(struct kobject *kobj,
+                              struct kobj_attribute *attr,
+                              char *buf)
+{
+       return sprintf(buf, "%u\n", khugepaged_full_scans);
+}
+static struct kobj_attribute full_scans_attr =
+       __ATTR_RO(full_scans);
+
+static ssize_t khugepaged_defrag_show(struct kobject *kobj,
+                                     struct kobj_attribute *attr, char *buf)
+{
+       return single_flag_show(kobj, attr, buf,
+                               TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
+}
+static ssize_t khugepaged_defrag_store(struct kobject *kobj,
+                                      struct kobj_attribute *attr,
+                                      const char *buf, size_t count)
+{
+       return single_flag_store(kobj, attr, buf, count,
+                                TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
+}
+static struct kobj_attribute khugepaged_defrag_attr =
+       __ATTR(defrag, 0644, khugepaged_defrag_show,
+              khugepaged_defrag_store);
+
+/*
+ * max_ptes_none controls if khugepaged should collapse hugepages over
+ * any unmapped ptes in turn potentially increasing the memory
+ * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
+ * reduce the available free memory in the system as it
+ * runs. Increasing max_ptes_none will instead potentially reduce the
+ * free memory in the system during the khugepaged scan.
+ */
+static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
+                                            struct kobj_attribute *attr,
+                                            char *buf)
+{
+       return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
+}
+static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
+                                             struct kobj_attribute *attr,
+                                             const char *buf, size_t count)
+{
+       int err;
+       unsigned long max_ptes_none;
+
+       err = strict_strtoul(buf, 10, &max_ptes_none);
+       if (err || max_ptes_none > HPAGE_PMD_NR-1)
+               return -EINVAL;
+
+       khugepaged_max_ptes_none = max_ptes_none;
+
+       return count;
+}
+static struct kobj_attribute khugepaged_max_ptes_none_attr =
+       __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
+              khugepaged_max_ptes_none_store);
+
+static struct attribute *khugepaged_attr[] = {
+       &khugepaged_defrag_attr.attr,
+       &khugepaged_max_ptes_none_attr.attr,
+       &pages_to_scan_attr.attr,
+       &pages_collapsed_attr.attr,
+       &full_scans_attr.attr,
+       &scan_sleep_millisecs_attr.attr,
+       &alloc_sleep_millisecs_attr.attr,
+       NULL,
+};
+
+static struct attribute_group khugepaged_attr_group = {
+       .attrs = khugepaged_attr,
+       .name = "khugepaged",
 };
 #endif /* CONFIG_SYSFS */
 
 static int __init hugepage_init(void)
 {
-#ifdef CONFIG_SYSFS
        int err;
+#ifdef CONFIG_SYSFS
+       static struct kobject *hugepage_kobj;
+#endif
 
-       err = sysfs_create_group(mm_kobj, &hugepage_attr_group);
-       if (err)
-               printk(KERN_ERR "hugepage: register sysfs failed\n");
+       err = -EINVAL;
+       if (!has_transparent_hugepage()) {
+               transparent_hugepage_flags = 0;
+               goto out;
+       }
+
+#ifdef CONFIG_SYSFS
+       err = -ENOMEM;
+       hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
+       if (unlikely(!hugepage_kobj)) {
+               printk(KERN_ERR "hugepage: failed kobject create\n");
+               goto out;
+       }
+
+       err = sysfs_create_group(hugepage_kobj, &hugepage_attr_group);
+       if (err) {
+               printk(KERN_ERR "hugepage: failed register hugeage group\n");
+               goto out;
+       }
+
+       err = sysfs_create_group(hugepage_kobj, &khugepaged_attr_group);
+       if (err) {
+               printk(KERN_ERR "hugepage: failed register hugeage group\n");
+               goto out;
+       }
 #endif
-       return 0;
+
+       err = khugepaged_slab_init();
+       if (err)
+               goto out;
+
+       err = mm_slots_hash_init();
+       if (err) {
+               khugepaged_slab_free();
+               goto out;
+       }
+
+       /*
+        * By default disable transparent hugepages on smaller systems,
+        * where the extra memory used could hurt more than TLB overhead
+        * is likely to save.  The admin can still enable it through /sys.
+        */
+       if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
+               transparent_hugepage_flags = 0;
+
+       start_khugepaged();
+
+       set_recommended_min_free_kbytes();
+
+out:
+       return err;
 }
 module_init(hugepage_init)
 
@@ -268,11 +647,27 @@ static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
        return ret;
 }
 
+static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
+{
+       return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
+}
+
+static inline struct page *alloc_hugepage_vma(int defrag,
+                                             struct vm_area_struct *vma,
+                                             unsigned long haddr, int nd,
+                                             gfp_t extra_gfp)
+{
+       return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
+                              HPAGE_PMD_ORDER, vma, haddr, nd);
+}
+
+#ifndef CONFIG_NUMA
 static inline struct page *alloc_hugepage(int defrag)
 {
-       return alloc_pages(GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT),
+       return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
                           HPAGE_PMD_ORDER);
 }
+#endif
 
 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
                               unsigned long address, pmd_t *pmd,
@@ -285,9 +680,15 @@ int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
        if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
                if (unlikely(anon_vma_prepare(vma)))
                        return VM_FAULT_OOM;
-               page = alloc_hugepage(transparent_hugepage_defrag(vma));
-               if (unlikely(!page))
+               if (unlikely(khugepaged_enter(vma)))
+                       return VM_FAULT_OOM;
+               page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
+                                         vma, haddr, numa_node_id(), 0);
+               if (unlikely(!page)) {
+                       count_vm_event(THP_FAULT_FALLBACK);
                        goto out;
+               }
+               count_vm_event(THP_FAULT_ALLOC);
                if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
                        put_page(page);
                        goto out;
@@ -406,8 +807,9 @@ static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
        }
 
        for (i = 0; i < HPAGE_PMD_NR; i++) {
-               pages[i] = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
-                                         vma, address);
+               pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
+                                              __GFP_OTHER_NODE,
+                                              vma, address, page_to_nid(page));
                if (unlikely(!pages[i] ||
                             mem_cgroup_newpage_charge(pages[i], mm,
                                                       GFP_KERNEL))) {
@@ -508,16 +910,19 @@ int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
 
        if (transparent_hugepage_enabled(vma) &&
            !transparent_hugepage_debug_cow())
-               new_page = alloc_hugepage(transparent_hugepage_defrag(vma));
+               new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
+                                             vma, haddr, numa_node_id(), 0);
        else
                new_page = NULL;
 
        if (unlikely(!new_page)) {
+               count_vm_event(THP_FAULT_FALLBACK);
                ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
                                                   pmd, orig_pmd, page, haddr);
                put_page(page);
                goto out;
        }
+       count_vm_event(THP_FAULT_ALLOC);
 
        if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
                put_page(new_page);
@@ -622,6 +1027,58 @@ int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
        return ret;
 }
 
+int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
+               unsigned long addr, unsigned long end,
+               unsigned char *vec)
+{
+       int ret = 0;
+
+       spin_lock(&vma->vm_mm->page_table_lock);
+       if (likely(pmd_trans_huge(*pmd))) {
+               ret = !pmd_trans_splitting(*pmd);
+               spin_unlock(&vma->vm_mm->page_table_lock);
+               if (unlikely(!ret))
+                       wait_split_huge_page(vma->anon_vma, pmd);
+               else {
+                       /*
+                        * All logical pages in the range are present
+                        * if backed by a huge page.
+                        */
+                       memset(vec, 1, (end - addr) >> PAGE_SHIFT);
+               }
+       } else
+               spin_unlock(&vma->vm_mm->page_table_lock);
+
+       return ret;
+}
+
+int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
+               unsigned long addr, pgprot_t newprot)
+{
+       struct mm_struct *mm = vma->vm_mm;
+       int ret = 0;
+
+       spin_lock(&mm->page_table_lock);
+       if (likely(pmd_trans_huge(*pmd))) {
+               if (unlikely(pmd_trans_splitting(*pmd))) {
+                       spin_unlock(&mm->page_table_lock);
+                       wait_split_huge_page(vma->anon_vma, pmd);
+               } else {
+                       pmd_t entry;
+
+                       entry = pmdp_get_and_clear(mm, addr, pmd);
+                       entry = pmd_modify(entry, newprot);
+                       set_pmd_at(mm, addr, pmd, entry);
+                       spin_unlock(&vma->vm_mm->page_table_lock);
+                       flush_tlb_range(vma, addr, addr + HPAGE_PMD_SIZE);
+                       ret = 1;
+               }
+       } else
+               spin_unlock(&vma->vm_mm->page_table_lock);
+
+       return ret;
+}
+
 pmd_t *page_check_address_pmd(struct page *page,
                              struct mm_struct *mm,
                              unsigned long address,
@@ -647,8 +1104,16 @@ pmd_t *page_check_address_pmd(struct page *page,
                goto out;
        if (pmd_page(*pmd) != page)
                goto out;
-       VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
-                 pmd_trans_splitting(*pmd));
+       /*
+        * split_vma() may create temporary aliased mappings. There is
+        * no risk as long as all huge pmd are found and have their
+        * splitting bit set before __split_huge_page_refcount
+        * runs. Finding the same huge pmd more than once during the
+        * same rmap walk is not a problem.
+        */
+       if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
+           pmd_trans_splitting(*pmd))
+               goto out;
        if (pmd_trans_huge(*pmd)) {
                VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
                          !pmd_trans_splitting(*pmd));
@@ -674,7 +1139,7 @@ static int __split_huge_page_splitting(struct page *page,
                 * We can't temporarily set the pmd to null in order
                 * to split it, the pmd must remain marked huge at all
                 * times or the VM won't take the pmd_trans_huge paths
-                * and it won't wait on the anon_vma->root->lock to
+                * and it won't wait on the anon_vma->root->mutex to
                 * serialize against split_huge_page*.
                 */
                pmdp_splitting_flush_notify(vma, address, pmd);
@@ -690,6 +1155,7 @@ static void __split_huge_page_refcount(struct page *page)
        int i;
        unsigned long head_index = page->index;
        struct zone *zone = page_zone(page);
+       int zonestat;
 
        /* prevent PageLRU to go away from under us, and freeze lru stats */
        spin_lock_irq(&zone->lru_lock);
@@ -707,7 +1173,12 @@ static void __split_huge_page_refcount(struct page *page)
                /* after clearing PageTail the gup refcount can be released */
                smp_mb();
 
-               page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
+               /*
+                * retain hwpoison flag of the poisoned tail page:
+                *   fix for the unsuitable process killed on Guest Machine(KVM)
+                *   by the memory-failure.
+                */
+               page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
                page_tail->flags |= (page->flags &
                                     ((1L << PG_referenced) |
                                      (1L << PG_swapbacked) |
@@ -748,9 +1219,23 @@ static void __split_huge_page_refcount(struct page *page)
                BUG_ON(!PageDirty(page_tail));
                BUG_ON(!PageSwapBacked(page_tail));
 
+               mem_cgroup_split_huge_fixup(page, page_tail);
+
                lru_add_page_tail(zone, page, page_tail);
        }
 
+       __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
+       __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
+
+       /*
+        * A hugepage counts for HPAGE_PMD_NR pages on the LRU statistics,
+        * so adjust those appropriately if this page is on the LRU.
+        */
+       if (PageLRU(page)) {
+               zonestat = NR_LRU_BASE + page_lru(page);
+               __mod_zone_page_state(zone, zonestat, -(HPAGE_PMD_NR-1));
+       }
+
        ClearPageCompound(page);
        compound_unlock(page);
        spin_unlock_irq(&zone->lru_lock);
@@ -848,7 +1333,7 @@ static int __split_huge_page_map(struct page *page,
        return ret;
 }
 
-/* must be called with anon_vma->root->lock hold */
+/* must be called with anon_vma->root->mutex hold */
 static void __split_huge_page(struct page *page,
                              struct anon_vma *anon_vma)
 {
@@ -914,6 +1399,7 @@ int split_huge_page(struct page *page)
 
        BUG_ON(!PageSwapBacked(page));
        __split_huge_page(page, anon_vma);
+       count_vm_event(THP_SPLIT);
 
        BUG_ON(PageCompound(page));
 out_unlock:
@@ -922,18 +1408,894 @@ out:
        return ret;
 }
 
-int hugepage_madvise(unsigned long *vm_flags)
+#define VM_NO_THP (VM_SPECIAL|VM_INSERTPAGE|VM_MIXEDMAP|VM_SAO| \
+                  VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
+
+int hugepage_madvise(struct vm_area_struct *vma,
+                    unsigned long *vm_flags, int advice)
+{
+       switch (advice) {
+       case MADV_HUGEPAGE:
+               /*
+                * Be somewhat over-protective like KSM for now!
+                */
+               if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
+                       return -EINVAL;
+               *vm_flags &= ~VM_NOHUGEPAGE;
+               *vm_flags |= VM_HUGEPAGE;
+               /*
+                * If the vma become good for khugepaged to scan,
+                * register it here without waiting a page fault that
+                * may not happen any time soon.
+                */
+               if (unlikely(khugepaged_enter_vma_merge(vma)))
+                       return -ENOMEM;
+               break;
+       case MADV_NOHUGEPAGE:
+               /*
+                * Be somewhat over-protective like KSM for now!
+                */
+               if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
+                       return -EINVAL;
+               *vm_flags &= ~VM_HUGEPAGE;
+               *vm_flags |= VM_NOHUGEPAGE;
+               /*
+                * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
+                * this vma even if we leave the mm registered in khugepaged if
+                * it got registered before VM_NOHUGEPAGE was set.
+                */
+               break;
+       }
+
+       return 0;
+}
+
+static int __init khugepaged_slab_init(void)
+{
+       mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
+                                         sizeof(struct mm_slot),
+                                         __alignof__(struct mm_slot), 0, NULL);
+       if (!mm_slot_cache)
+               return -ENOMEM;
+
+       return 0;
+}
+
+static void __init khugepaged_slab_free(void)
+{
+       kmem_cache_destroy(mm_slot_cache);
+       mm_slot_cache = NULL;
+}
+
+static inline struct mm_slot *alloc_mm_slot(void)
 {
+       if (!mm_slot_cache)     /* initialization failed */
+               return NULL;
+       return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
+}
+
+static inline void free_mm_slot(struct mm_slot *mm_slot)
+{
+       kmem_cache_free(mm_slot_cache, mm_slot);
+}
+
+static int __init mm_slots_hash_init(void)
+{
+       mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
+                               GFP_KERNEL);
+       if (!mm_slots_hash)
+               return -ENOMEM;
+       return 0;
+}
+
+#if 0
+static void __init mm_slots_hash_free(void)
+{
+       kfree(mm_slots_hash);
+       mm_slots_hash = NULL;
+}
+#endif
+
+static struct mm_slot *get_mm_slot(struct mm_struct *mm)
+{
+       struct mm_slot *mm_slot;
+       struct hlist_head *bucket;
+       struct hlist_node *node;
+
+       bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
+                               % MM_SLOTS_HASH_HEADS];
+       hlist_for_each_entry(mm_slot, node, bucket, hash) {
+               if (mm == mm_slot->mm)
+                       return mm_slot;
+       }
+       return NULL;
+}
+
+static void insert_to_mm_slots_hash(struct mm_struct *mm,
+                                   struct mm_slot *mm_slot)
+{
+       struct hlist_head *bucket;
+
+       bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
+                               % MM_SLOTS_HASH_HEADS];
+       mm_slot->mm = mm;
+       hlist_add_head(&mm_slot->hash, bucket);
+}
+
+static inline int khugepaged_test_exit(struct mm_struct *mm)
+{
+       return atomic_read(&mm->mm_users) == 0;
+}
+
+int __khugepaged_enter(struct mm_struct *mm)
+{
+       struct mm_slot *mm_slot;
+       int wakeup;
+
+       mm_slot = alloc_mm_slot();
+       if (!mm_slot)
+               return -ENOMEM;
+
+       /* __khugepaged_exit() must not run from under us */
+       VM_BUG_ON(khugepaged_test_exit(mm));
+       if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
+               free_mm_slot(mm_slot);
+               return 0;
+       }
+
+       spin_lock(&khugepaged_mm_lock);
+       insert_to_mm_slots_hash(mm, mm_slot);
        /*
-        * Be somewhat over-protective like KSM for now!
+        * Insert just behind the scanning cursor, to let the area settle
+        * down a little.
         */
-       if (*vm_flags & (VM_HUGEPAGE | VM_SHARED  | VM_MAYSHARE   |
-                        VM_PFNMAP   | VM_IO      | VM_DONTEXPAND |
-                        VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
-                        VM_MIXEDMAP | VM_SAO))
-               return -EINVAL;
+       wakeup = list_empty(&khugepaged_scan.mm_head);
+       list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
+       spin_unlock(&khugepaged_mm_lock);
 
-       *vm_flags |= VM_HUGEPAGE;
+       atomic_inc(&mm->mm_count);
+       if (wakeup)
+               wake_up_interruptible(&khugepaged_wait);
+
+       return 0;
+}
+
+int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
+{
+       unsigned long hstart, hend;
+       if (!vma->anon_vma)
+               /*
+                * Not yet faulted in so we will register later in the
+                * page fault if needed.
+                */
+               return 0;
+       if (vma->vm_ops)
+               /* khugepaged not yet working on file or special mappings */
+               return 0;
+       /*
+        * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
+        * true too, verify it here.
+        */
+       VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
+       hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
+       hend = vma->vm_end & HPAGE_PMD_MASK;
+       if (hstart < hend)
+               return khugepaged_enter(vma);
+       return 0;
+}
+
+void __khugepaged_exit(struct mm_struct *mm)
+{
+       struct mm_slot *mm_slot;
+       int free = 0;
+
+       spin_lock(&khugepaged_mm_lock);
+       mm_slot = get_mm_slot(mm);
+       if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
+               hlist_del(&mm_slot->hash);
+               list_del(&mm_slot->mm_node);
+               free = 1;
+       }
+       spin_unlock(&khugepaged_mm_lock);
+
+       if (free) {
+               clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
+               free_mm_slot(mm_slot);
+               mmdrop(mm);
+       } else if (mm_slot) {
+               /*
+                * This is required to serialize against
+                * khugepaged_test_exit() (which is guaranteed to run
+                * under mmap sem read mode). Stop here (after we
+                * return all pagetables will be destroyed) until
+                * khugepaged has finished working on the pagetables
+                * under the mmap_sem.
+                */
+               down_write(&mm->mmap_sem);
+               up_write(&mm->mmap_sem);
+       }
+}
+
+static void release_pte_page(struct page *page)
+{
+       /* 0 stands for page_is_file_cache(page) == false */
+       dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
+       unlock_page(page);
+       putback_lru_page(page);
+}
+
+static void release_pte_pages(pte_t *pte, pte_t *_pte)
+{
+       while (--_pte >= pte) {
+               pte_t pteval = *_pte;
+               if (!pte_none(pteval))
+                       release_pte_page(pte_page(pteval));
+       }
+}
+
+static void release_all_pte_pages(pte_t *pte)
+{
+       release_pte_pages(pte, pte + HPAGE_PMD_NR);
+}
+
+static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
+                                       unsigned long address,
+                                       pte_t *pte)
+{
+       struct page *page;
+       pte_t *_pte;
+       int referenced = 0, isolated = 0, none = 0;
+       for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
+            _pte++, address += PAGE_SIZE) {
+               pte_t pteval = *_pte;
+               if (pte_none(pteval)) {
+                       if (++none <= khugepaged_max_ptes_none)
+                               continue;
+                       else {
+                               release_pte_pages(pte, _pte);
+                               goto out;
+                       }
+               }
+               if (!pte_present(pteval) || !pte_write(pteval)) {
+                       release_pte_pages(pte, _pte);
+                       goto out;
+               }
+               page = vm_normal_page(vma, address, pteval);
+               if (unlikely(!page)) {
+                       release_pte_pages(pte, _pte);
+                       goto out;
+               }
+               VM_BUG_ON(PageCompound(page));
+               BUG_ON(!PageAnon(page));
+               VM_BUG_ON(!PageSwapBacked(page));
+
+               /* cannot use mapcount: can't collapse if there's a gup pin */
+               if (page_count(page) != 1) {
+                       release_pte_pages(pte, _pte);
+                       goto out;
+               }
+               /*
+                * We can do it before isolate_lru_page because the
+                * page can't be freed from under us. NOTE: PG_lock
+                * is needed to serialize against split_huge_page
+                * when invoked from the VM.
+                */
+               if (!trylock_page(page)) {
+                       release_pte_pages(pte, _pte);
+                       goto out;
+               }
+               /*
+                * Isolate the page to avoid collapsing an hugepage
+                * currently in use by the VM.
+                */
+               if (isolate_lru_page(page)) {
+                       unlock_page(page);
+                       release_pte_pages(pte, _pte);
+                       goto out;
+               }
+               /* 0 stands for page_is_file_cache(page) == false */
+               inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
+               VM_BUG_ON(!PageLocked(page));
+               VM_BUG_ON(PageLRU(page));
+
+               /* If there is no mapped pte young don't collapse the page */
+               if (pte_young(pteval) || PageReferenced(page) ||
+                   mmu_notifier_test_young(vma->vm_mm, address))
+                       referenced = 1;
+       }
+       if (unlikely(!referenced))
+               release_all_pte_pages(pte);
+       else
+               isolated = 1;
+out:
+       return isolated;
+}
+
+static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
+                                     struct vm_area_struct *vma,
+                                     unsigned long address,
+                                     spinlock_t *ptl)
+{
+       pte_t *_pte;
+       for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
+               pte_t pteval = *_pte;
+               struct page *src_page;
+
+               if (pte_none(pteval)) {
+                       clear_user_highpage(page, address);
+                       add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
+               } else {
+                       src_page = pte_page(pteval);
+                       copy_user_highpage(page, src_page, address, vma);
+                       VM_BUG_ON(page_mapcount(src_page) != 1);
+                       VM_BUG_ON(page_count(src_page) != 2);
+                       release_pte_page(src_page);
+                       /*
+                        * ptl mostly unnecessary, but preempt has to
+                        * be disabled to update the per-cpu stats
+                        * inside page_remove_rmap().
+                        */
+                       spin_lock(ptl);
+                       /*
+                        * paravirt calls inside pte_clear here are
+                        * superfluous.
+                        */
+                       pte_clear(vma->vm_mm, address, _pte);
+                       page_remove_rmap(src_page);
+                       spin_unlock(ptl);
+                       free_page_and_swap_cache(src_page);
+               }
+
+               address += PAGE_SIZE;
+               page++;
+       }
+}
+
+static void collapse_huge_page(struct mm_struct *mm,
+                              unsigned long address,
+                              struct page **hpage,
+                              struct vm_area_struct *vma,
+                              int node)
+{
+       pgd_t *pgd;
+       pud_t *pud;
+       pmd_t *pmd, _pmd;
+       pte_t *pte;
+       pgtable_t pgtable;
+       struct page *new_page;
+       spinlock_t *ptl;
+       int isolated;
+       unsigned long hstart, hend;
+
+       VM_BUG_ON(address & ~HPAGE_PMD_MASK);
+#ifndef CONFIG_NUMA
+       up_read(&mm->mmap_sem);
+       VM_BUG_ON(!*hpage);
+       new_page = *hpage;
+#else
+       VM_BUG_ON(*hpage);
+       /*
+        * Allocate the page while the vma is still valid and under
+        * the mmap_sem read mode so there is no memory allocation
+        * later when we take the mmap_sem in write mode. This is more
+        * friendly behavior (OTOH it may actually hide bugs) to
+        * filesystems in userland with daemons allocating memory in
+        * the userland I/O paths.  Allocating memory with the
+        * mmap_sem in read mode is good idea also to allow greater
+        * scalability.
+        */
+       new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
+                                     node, __GFP_OTHER_NODE);
+
+       /*
+        * After allocating the hugepage, release the mmap_sem read lock in
+        * preparation for taking it in write mode.
+        */
+       up_read(&mm->mmap_sem);
+       if (unlikely(!new_page)) {
+               count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
+               *hpage = ERR_PTR(-ENOMEM);
+               return;
+       }
+#endif
+
+       count_vm_event(THP_COLLAPSE_ALLOC);
+       if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
+#ifdef CONFIG_NUMA
+               put_page(new_page);
+#endif
+               return;
+       }
+
+       /*
+        * Prevent all access to pagetables with the exception of
+        * gup_fast later hanlded by the ptep_clear_flush and the VM
+        * handled by the anon_vma lock + PG_lock.
+        */
+       down_write(&mm->mmap_sem);
+       if (unlikely(khugepaged_test_exit(mm)))
+               goto out;
+
+       vma = find_vma(mm, address);
+       hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
+       hend = vma->vm_end & HPAGE_PMD_MASK;
+       if (address < hstart || address + HPAGE_PMD_SIZE > hend)
+               goto out;
+
+       if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
+           (vma->vm_flags & VM_NOHUGEPAGE))
+               goto out;
+
+       if (!vma->anon_vma || vma->vm_ops)
+               goto out;
+       if (is_vma_temporary_stack(vma))
+               goto out;
+       /*
+        * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
+        * true too, verify it here.
+        */
+       VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
+
+       pgd = pgd_offset(mm, address);
+       if (!pgd_present(*pgd))
+               goto out;
+
+       pud = pud_offset(pgd, address);
+       if (!pud_present(*pud))
+               goto out;
+
+       pmd = pmd_offset(pud, address);
+       /* pmd can't go away or become huge under us */
+       if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
+               goto out;
+
+       anon_vma_lock(vma->anon_vma);
+
+       pte = pte_offset_map(pmd, address);
+       ptl = pte_lockptr(mm, pmd);
+
+       spin_lock(&mm->page_table_lock); /* probably unnecessary */
+       /*
+        * After this gup_fast can't run anymore. This also removes
+        * any huge TLB entry from the CPU so we won't allow
+        * huge and small TLB entries for the same virtual address
+        * to avoid the risk of CPU bugs in that area.
+        */
+       _pmd = pmdp_clear_flush_notify(vma, address, pmd);
+       spin_unlock(&mm->page_table_lock);
+
+       spin_lock(ptl);
+       isolated = __collapse_huge_page_isolate(vma, address, pte);
+       spin_unlock(ptl);
+
+       if (unlikely(!isolated)) {
+               pte_unmap(pte);
+               spin_lock(&mm->page_table_lock);
+               BUG_ON(!pmd_none(*pmd));
+               set_pmd_at(mm, address, pmd, _pmd);
+               spin_unlock(&mm->page_table_lock);
+               anon_vma_unlock(vma->anon_vma);
+               goto out;
+       }
+
+       /*
+        * All pages are isolated and locked so anon_vma rmap
+        * can't run anymore.
+        */
+       anon_vma_unlock(vma->anon_vma);
+
+       __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
+       pte_unmap(pte);
+       __SetPageUptodate(new_page);
+       pgtable = pmd_pgtable(_pmd);
+       VM_BUG_ON(page_count(pgtable) != 1);
+       VM_BUG_ON(page_mapcount(pgtable) != 0);
+
+       _pmd = mk_pmd(new_page, vma->vm_page_prot);
+       _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
+       _pmd = pmd_mkhuge(_pmd);
+
+       /*
+        * spin_lock() below is not the equivalent of smp_wmb(), so
+        * this is needed to avoid the copy_huge_page writes to become
+        * visible after the set_pmd_at() write.
+        */
+       smp_wmb();
+
+       spin_lock(&mm->page_table_lock);
+       BUG_ON(!pmd_none(*pmd));
+       page_add_new_anon_rmap(new_page, vma, address);
+       set_pmd_at(mm, address, pmd, _pmd);
+       update_mmu_cache(vma, address, entry);
+       prepare_pmd_huge_pte(pgtable, mm);
+       mm->nr_ptes--;
+       spin_unlock(&mm->page_table_lock);
+
+#ifndef CONFIG_NUMA
+       *hpage = NULL;
+#endif
+       khugepaged_pages_collapsed++;
+out_up_write:
+       up_write(&mm->mmap_sem);
+       return;
+
+out:
+       mem_cgroup_uncharge_page(new_page);
+#ifdef CONFIG_NUMA
+       put_page(new_page);
+#endif
+       goto out_up_write;
+}
+
+static int khugepaged_scan_pmd(struct mm_struct *mm,
+                              struct vm_area_struct *vma,
+                              unsigned long address,
+                              struct page **hpage)
+{
+       pgd_t *pgd;
+       pud_t *pud;
+       pmd_t *pmd;
+       pte_t *pte, *_pte;
+       int ret = 0, referenced = 0, none = 0;
+       struct page *page;
+       unsigned long _address;
+       spinlock_t *ptl;
+       int node = -1;
+
+       VM_BUG_ON(address & ~HPAGE_PMD_MASK);
+
+       pgd = pgd_offset(mm, address);
+       if (!pgd_present(*pgd))
+               goto out;
+
+       pud = pud_offset(pgd, address);
+       if (!pud_present(*pud))
+               goto out;
+
+       pmd = pmd_offset(pud, address);
+       if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
+               goto out;
+
+       pte = pte_offset_map_lock(mm, pmd, address, &ptl);
+       for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
+            _pte++, _address += PAGE_SIZE) {
+               pte_t pteval = *_pte;
+               if (pte_none(pteval)) {
+                       if (++none <= khugepaged_max_ptes_none)
+                               continue;
+                       else
+                               goto out_unmap;
+               }
+               if (!pte_present(pteval) || !pte_write(pteval))
+                       goto out_unmap;
+               page = vm_normal_page(vma, _address, pteval);
+               if (unlikely(!page))
+                       goto out_unmap;
+               /*
+                * Chose the node of the first page. This could
+                * be more sophisticated and look at more pages,
+                * but isn't for now.
+                */
+               if (node == -1)
+                       node = page_to_nid(page);
+               VM_BUG_ON(PageCompound(page));
+               if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
+                       goto out_unmap;
+               /* cannot use mapcount: can't collapse if there's a gup pin */
+               if (page_count(page) != 1)
+                       goto out_unmap;
+               if (pte_young(pteval) || PageReferenced(page) ||
+                   mmu_notifier_test_young(vma->vm_mm, address))
+                       referenced = 1;
+       }
+       if (referenced)
+               ret = 1;
+out_unmap:
+       pte_unmap_unlock(pte, ptl);
+       if (ret)
+               /* collapse_huge_page will return with the mmap_sem released */
+               collapse_huge_page(mm, address, hpage, vma, node);
+out:
+       return ret;
+}
+
+static void collect_mm_slot(struct mm_slot *mm_slot)
+{
+       struct mm_struct *mm = mm_slot->mm;
+
+       VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock));
+
+       if (khugepaged_test_exit(mm)) {
+               /* free mm_slot */
+               hlist_del(&mm_slot->hash);
+               list_del(&mm_slot->mm_node);
+
+               /*
+                * Not strictly needed because the mm exited already.
+                *
+                * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
+                */
+
+               /* khugepaged_mm_lock actually not necessary for the below */
+               free_mm_slot(mm_slot);
+               mmdrop(mm);
+       }
+}
+
+static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
+                                           struct page **hpage)
+{
+       struct mm_slot *mm_slot;
+       struct mm_struct *mm;
+       struct vm_area_struct *vma;
+       int progress = 0;
+
+       VM_BUG_ON(!pages);
+       VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock));
+
+       if (khugepaged_scan.mm_slot)
+               mm_slot = khugepaged_scan.mm_slot;
+       else {
+               mm_slot = list_entry(khugepaged_scan.mm_head.next,
+                                    struct mm_slot, mm_node);
+               khugepaged_scan.address = 0;
+               khugepaged_scan.mm_slot = mm_slot;
+       }
+       spin_unlock(&khugepaged_mm_lock);
+
+       mm = mm_slot->mm;
+       down_read(&mm->mmap_sem);
+       if (unlikely(khugepaged_test_exit(mm)))
+               vma = NULL;
+       else
+               vma = find_vma(mm, khugepaged_scan.address);
+
+       progress++;
+       for (; vma; vma = vma->vm_next) {
+               unsigned long hstart, hend;
+
+               cond_resched();
+               if (unlikely(khugepaged_test_exit(mm))) {
+                       progress++;
+                       break;
+               }
+
+               if ((!(vma->vm_flags & VM_HUGEPAGE) &&
+                    !khugepaged_always()) ||
+                   (vma->vm_flags & VM_NOHUGEPAGE)) {
+               skip:
+                       progress++;
+                       continue;
+               }
+               if (!vma->anon_vma || vma->vm_ops)
+                       goto skip;
+               if (is_vma_temporary_stack(vma))
+                       goto skip;
+               /*
+                * If is_pfn_mapping() is true is_learn_pfn_mapping()
+                * must be true too, verify it here.
+                */
+               VM_BUG_ON(is_linear_pfn_mapping(vma) ||
+                         vma->vm_flags & VM_NO_THP);
+
+               hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
+               hend = vma->vm_end & HPAGE_PMD_MASK;
+               if (hstart >= hend)
+                       goto skip;
+               if (khugepaged_scan.address > hend)
+                       goto skip;
+               if (khugepaged_scan.address < hstart)
+                       khugepaged_scan.address = hstart;
+               VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
+
+               while (khugepaged_scan.address < hend) {
+                       int ret;
+                       cond_resched();
+                       if (unlikely(khugepaged_test_exit(mm)))
+                               goto breakouterloop;
+
+                       VM_BUG_ON(khugepaged_scan.address < hstart ||
+                                 khugepaged_scan.address + HPAGE_PMD_SIZE >
+                                 hend);
+                       ret = khugepaged_scan_pmd(mm, vma,
+                                                 khugepaged_scan.address,
+                                                 hpage);
+                       /* move to next address */
+                       khugepaged_scan.address += HPAGE_PMD_SIZE;
+                       progress += HPAGE_PMD_NR;
+                       if (ret)
+                               /* we released mmap_sem so break loop */
+                               goto breakouterloop_mmap_sem;
+                       if (progress >= pages)
+                               goto breakouterloop;
+               }
+       }
+breakouterloop:
+       up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
+breakouterloop_mmap_sem:
+
+       spin_lock(&khugepaged_mm_lock);
+       VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
+       /*
+        * Release the current mm_slot if this mm is about to die, or
+        * if we scanned all vmas of this mm.
+        */
+       if (khugepaged_test_exit(mm) || !vma) {
+               /*
+                * Make sure that if mm_users is reaching zero while
+                * khugepaged runs here, khugepaged_exit will find
+                * mm_slot not pointing to the exiting mm.
+                */
+               if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
+                       khugepaged_scan.mm_slot = list_entry(
+                               mm_slot->mm_node.next,
+                               struct mm_slot, mm_node);
+                       khugepaged_scan.address = 0;
+               } else {
+                       khugepaged_scan.mm_slot = NULL;
+                       khugepaged_full_scans++;
+               }
+
+               collect_mm_slot(mm_slot);
+       }
+
+       return progress;
+}
+
+static int khugepaged_has_work(void)
+{
+       return !list_empty(&khugepaged_scan.mm_head) &&
+               khugepaged_enabled();
+}
+
+static int khugepaged_wait_event(void)
+{
+       return !list_empty(&khugepaged_scan.mm_head) ||
+               !khugepaged_enabled();
+}
+
+static void khugepaged_do_scan(struct page **hpage)
+{
+       unsigned int progress = 0, pass_through_head = 0;
+       unsigned int pages = khugepaged_pages_to_scan;
+
+       barrier(); /* write khugepaged_pages_to_scan to local stack */
+
+       while (progress < pages) {
+               cond_resched();
+
+#ifndef CONFIG_NUMA
+               if (!*hpage) {
+                       *hpage = alloc_hugepage(khugepaged_defrag());
+                       if (unlikely(!*hpage)) {
+                               count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
+                               break;
+                       }
+                       count_vm_event(THP_COLLAPSE_ALLOC);
+               }
+#else
+               if (IS_ERR(*hpage))
+                       break;
+#endif
+
+               if (unlikely(kthread_should_stop() || freezing(current)))
+                       break;
+
+               spin_lock(&khugepaged_mm_lock);
+               if (!khugepaged_scan.mm_slot)
+                       pass_through_head++;
+               if (khugepaged_has_work() &&
+                   pass_through_head < 2)
+                       progress += khugepaged_scan_mm_slot(pages - progress,
+                                                           hpage);
+               else
+                       progress = pages;
+               spin_unlock(&khugepaged_mm_lock);
+       }
+}
+
+static void khugepaged_alloc_sleep(void)
+{
+       DEFINE_WAIT(wait);
+       add_wait_queue(&khugepaged_wait, &wait);
+       schedule_timeout_interruptible(
+               msecs_to_jiffies(
+                       khugepaged_alloc_sleep_millisecs));
+       remove_wait_queue(&khugepaged_wait, &wait);
+}
+
+#ifndef CONFIG_NUMA
+static struct page *khugepaged_alloc_hugepage(void)
+{
+       struct page *hpage;
+
+       do {
+               hpage = alloc_hugepage(khugepaged_defrag());
+               if (!hpage) {
+                       count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
+                       khugepaged_alloc_sleep();
+               } else
+                       count_vm_event(THP_COLLAPSE_ALLOC);
+       } while (unlikely(!hpage) &&
+                likely(khugepaged_enabled()));
+       return hpage;
+}
+#endif
+
+static void khugepaged_loop(void)
+{
+       struct page *hpage;
+
+#ifdef CONFIG_NUMA
+       hpage = NULL;
+#endif
+       while (likely(khugepaged_enabled())) {
+#ifndef CONFIG_NUMA
+               hpage = khugepaged_alloc_hugepage();
+               if (unlikely(!hpage))
+                       break;
+#else
+               if (IS_ERR(hpage)) {
+                       khugepaged_alloc_sleep();
+                       hpage = NULL;
+               }
+#endif
+
+               khugepaged_do_scan(&hpage);
+#ifndef CONFIG_NUMA
+               if (hpage)
+                       put_page(hpage);
+#endif
+               try_to_freeze();
+               if (unlikely(kthread_should_stop()))
+                       break;
+               if (khugepaged_has_work()) {
+                       DEFINE_WAIT(wait);
+                       if (!khugepaged_scan_sleep_millisecs)
+                               continue;
+                       add_wait_queue(&khugepaged_wait, &wait);
+                       schedule_timeout_interruptible(
+                               msecs_to_jiffies(
+                                       khugepaged_scan_sleep_millisecs));
+                       remove_wait_queue(&khugepaged_wait, &wait);
+               } else if (khugepaged_enabled())
+                       wait_event_freezable(khugepaged_wait,
+                                            khugepaged_wait_event());
+       }
+}
+
+static int khugepaged(void *none)
+{
+       struct mm_slot *mm_slot;
+
+       set_freezable();
+       set_user_nice(current, 19);
+
+       /* serialize with start_khugepaged() */
+       mutex_lock(&khugepaged_mutex);
+
+       for (;;) {
+               mutex_unlock(&khugepaged_mutex);
+               VM_BUG_ON(khugepaged_thread != current);
+               khugepaged_loop();
+               VM_BUG_ON(khugepaged_thread != current);
+
+               mutex_lock(&khugepaged_mutex);
+               if (!khugepaged_enabled())
+                       break;
+               if (unlikely(kthread_should_stop()))
+                       break;
+       }
+
+       spin_lock(&khugepaged_mm_lock);
+       mm_slot = khugepaged_scan.mm_slot;
+       khugepaged_scan.mm_slot = NULL;
+       if (mm_slot)
+               collect_mm_slot(mm_slot);
+       spin_unlock(&khugepaged_mm_lock);
+
+       khugepaged_thread = NULL;
+       mutex_unlock(&khugepaged_mutex);
 
        return 0;
 }
@@ -957,3 +2319,71 @@ void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
        put_page(page);
        BUG_ON(pmd_trans_huge(*pmd));
 }
+
+static void split_huge_page_address(struct mm_struct *mm,
+                                   unsigned long address)
+{
+       pgd_t *pgd;
+       pud_t *pud;
+       pmd_t *pmd;
+
+       VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
+
+       pgd = pgd_offset(mm, address);
+       if (!pgd_present(*pgd))
+               return;
+
+       pud = pud_offset(pgd, address);
+       if (!pud_present(*pud))
+               return;
+
+       pmd = pmd_offset(pud, address);
+       if (!pmd_present(*pmd))
+               return;
+       /*
+        * Caller holds the mmap_sem write mode, so a huge pmd cannot
+        * materialize from under us.
+        */
+       split_huge_page_pmd(mm, pmd);
+}
+
+void __vma_adjust_trans_huge(struct vm_area_struct *vma,
+                            unsigned long start,
+                            unsigned long end,
+                            long adjust_next)
+{
+       /*
+        * If the new start address isn't hpage aligned and it could
+        * previously contain an hugepage: check if we need to split
+        * an huge pmd.
+        */
+       if (start & ~HPAGE_PMD_MASK &&
+           (start & HPAGE_PMD_MASK) >= vma->vm_start &&
+           (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
+               split_huge_page_address(vma->vm_mm, start);
+
+       /*
+        * If the new end address isn't hpage aligned and it could
+        * previously contain an hugepage: check if we need to split
+        * an huge pmd.
+        */
+       if (end & ~HPAGE_PMD_MASK &&
+           (end & HPAGE_PMD_MASK) >= vma->vm_start &&
+           (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
+               split_huge_page_address(vma->vm_mm, end);
+
+       /*
+        * If we're also updating the vma->vm_next->vm_start, if the new
+        * vm_next->vm_start isn't page aligned and it could previously
+        * contain an hugepage: check if we need to split an huge pmd.
+        */
+       if (adjust_next > 0) {
+               struct vm_area_struct *next = vma->vm_next;
+               unsigned long nstart = next->vm_start;
+               nstart += adjust_next << PAGE_SHIFT;
+               if (nstart & ~HPAGE_PMD_MASK &&
+                   (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
+                   (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
+                       split_huge_page_address(next->vm_mm, nstart);
+       }
+}