ksm: Kernel SamePage Merging
authorIzik Eidus <ieidus@redhat.com>
Tue, 22 Sep 2009 00:02:03 +0000 (17:02 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 22 Sep 2009 14:17:31 +0000 (07:17 -0700)
Ksm is code that allows merging of identical pages between one or more
applications, in a way invisible to the applications that use it.  Pages
that are merged are marked as read-only, then COWed when any application
tries to change them.

Whereas fork() allows sharing anonymous pages between parent and child,
ksm can share anonymous pages between unrelated processes.

Ksm works by walking over the memory pages of the applications it scans,
in order to find identical pages.  It uses two sorted data structures,
called the stable and unstable trees, to locate identical pages in an
effective way.

When ksm finds two identical pages, it marks them as readonly and merges
them into a single page.  After the pages have been marked as readonly and
merged into one, Linux treats them as normal copy-on-write pages, copying
to a fresh anonymous page if write access is required later.

Ksm scans and merges anonymous pages only in those memory areas that have
been registered with it by madvise(addr, length, MADV_MERGEABLE).

The ksm scanner is controlled by sysfs files in /sys/kernel/mm/ksm/:

max_kernel_pages - the maximum number of unswappable kernel pages
                   which may be allocated by ksm (0 for unlimited).

kernel_pages_allocated - how many ksm pages are currently allocated,
                         sharing identical content between different
                         processes (pages unswappable in this release).

pages_shared - how many pages have been saved by sharing with ksm pages
               (kernel_pages_allocated being excluded from this count).

pages_to_scan - how many pages ksm should scan before sleeping.

sleep_millisecs - how many milliseconds ksm should sleep between scans.

run - write 0 to disable ksm, read 0 while ksm is disabled (default),
      write 1 to run ksm, read 1 while ksm is running,
      write 2 to disable ksm and unmerge all its pages.

Includes contributions by Andrea Arcangeli Chris Wright and Hugh Dickins.

[hugh.dickins@tiscali.co.uk: fix rare page leak]
Signed-off-by: Izik Eidus <ieidus@redhat.com>
Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Chris Wright <chrisw@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Balbir Singh <balbir@in.ibm.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Avi Kivity <avi@redhat.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/ksm.c

index 8b76008fcd3240f5c8aa7a07f1b37d7756bfd101..2c02094807e0bd11abd98639bdd6691f492425ab 100644 (file)
--- a/mm/ksm.c
+++ b/mm/ksm.c
 /*
- * Initial dummy version just to illustrate KSM's interface to other files.
+ * Memory merging support.
+ *
+ * This code enables dynamic sharing of identical pages found in different
+ * memory areas, even if they are not shared by fork()
+ *
+ * Copyright (C) 2008 Red Hat, Inc.
+ * Authors:
+ *     Izik Eidus
+ *     Andrea Arcangeli
+ *     Chris Wright
+ *
+ * This work is licensed under the terms of the GNU GPL, version 2.
  */
 
 #include <linux/errno.h>
+#include <linux/mm.h>
+#include <linux/fs.h>
 #include <linux/mman.h>
+#include <linux/sched.h>
+#include <linux/rwsem.h>
+#include <linux/pagemap.h>
+#include <linux/rmap.h>
+#include <linux/spinlock.h>
+#include <linux/jhash.h>
+#include <linux/delay.h>
+#include <linux/kthread.h>
+#include <linux/wait.h>
+#include <linux/slab.h>
+#include <linux/rbtree.h>
+#include <linux/mmu_notifier.h>
 #include <linux/ksm.h>
 
+#include <asm/tlbflush.h>
+
+/*
+ * A few notes about the KSM scanning process,
+ * to make it easier to understand the data structures below:
+ *
+ * In order to reduce excessive scanning, KSM sorts the memory pages by their
+ * contents into a data structure that holds pointers to the pages' locations.
+ *
+ * Since the contents of the pages may change at any moment, KSM cannot just
+ * insert the pages into a normal sorted tree and expect it to find anything.
+ * Therefore KSM uses two data structures - the stable and the unstable tree.
+ *
+ * The stable tree holds pointers to all the merged pages (ksm pages), sorted
+ * by their contents.  Because each such page is write-protected, searching on
+ * this tree is fully assured to be working (except when pages are unmapped),
+ * and therefore this tree is called the stable tree.
+ *
+ * In addition to the stable tree, KSM uses a second data structure called the
+ * unstable tree: this tree holds pointers to pages which have been found to
+ * be "unchanged for a period of time".  The unstable tree sorts these pages
+ * by their contents, but since they are not write-protected, KSM cannot rely
+ * upon the unstable tree to work correctly - the unstable tree is liable to
+ * be corrupted as its contents are modified, and so it is called unstable.
+ *
+ * KSM solves this problem by several techniques:
+ *
+ * 1) The unstable tree is flushed every time KSM completes scanning all
+ *    memory areas, and then the tree is rebuilt again from the beginning.
+ * 2) KSM will only insert into the unstable tree, pages whose hash value
+ *    has not changed since the previous scan of all memory areas.
+ * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
+ *    colors of the nodes and not on their contents, assuring that even when
+ *    the tree gets "corrupted" it won't get out of balance, so scanning time
+ *    remains the same (also, searching and inserting nodes in an rbtree uses
+ *    the same algorithm, so we have no overhead when we flush and rebuild).
+ * 4) KSM never flushes the stable tree, which means that even if it were to
+ *    take 10 attempts to find a page in the unstable tree, once it is found,
+ *    it is secured in the stable tree.  (When we scan a new page, we first
+ *    compare it against the stable tree, and then against the unstable tree.)
+ */
+
+/**
+ * struct mm_slot - ksm information per mm that is being scanned
+ * @link: link to the mm_slots hash list
+ * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
+ * @rmap_list: head for this mm_slot's list of rmap_items
+ * @mm: the mm that this information is valid for
+ */
+struct mm_slot {
+       struct hlist_node link;
+       struct list_head mm_list;
+       struct list_head rmap_list;
+       struct mm_struct *mm;
+};
+
+/**
+ * struct ksm_scan - cursor for scanning
+ * @mm_slot: the current mm_slot we are scanning
+ * @address: the next address inside that to be scanned
+ * @rmap_item: the current rmap that we are scanning inside the rmap_list
+ * @seqnr: count of completed full scans (needed when removing unstable node)
+ *
+ * There is only the one ksm_scan instance of this cursor structure.
+ */
+struct ksm_scan {
+       struct mm_slot *mm_slot;
+       unsigned long address;
+       struct rmap_item *rmap_item;
+       unsigned long seqnr;
+};
+
+/**
+ * struct rmap_item - reverse mapping item for virtual addresses
+ * @link: link into mm_slot's rmap_list (rmap_list is per mm)
+ * @mm: the memory structure this rmap_item is pointing into
+ * @address: the virtual address this rmap_item tracks (+ flags in low bits)
+ * @oldchecksum: previous checksum of the page at that virtual address
+ * @node: rb_node of this rmap_item in either unstable or stable tree
+ * @next: next rmap_item hanging off the same node of the stable tree
+ * @prev: previous rmap_item hanging off the same node of the stable tree
+ */
+struct rmap_item {
+       struct list_head link;
+       struct mm_struct *mm;
+       unsigned long address;          /* + low bits used for flags below */
+       union {
+               unsigned int oldchecksum;               /* when unstable */
+               struct rmap_item *next;                 /* when stable */
+       };
+       union {
+               struct rb_node node;                    /* when tree node */
+               struct rmap_item *prev;                 /* in stable list */
+       };
+};
+
+#define SEQNR_MASK     0x0ff   /* low bits of unstable tree seqnr */
+#define NODE_FLAG      0x100   /* is a node of unstable or stable tree */
+#define STABLE_FLAG    0x200   /* is a node or list item of stable tree */
+
+/* The stable and unstable tree heads */
+static struct rb_root root_stable_tree = RB_ROOT;
+static struct rb_root root_unstable_tree = RB_ROOT;
+
+#define MM_SLOTS_HASH_HEADS 1024
+static struct hlist_head *mm_slots_hash;
+
+static struct mm_slot ksm_mm_head = {
+       .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
+};
+static struct ksm_scan ksm_scan = {
+       .mm_slot = &ksm_mm_head,
+};
+
+static struct kmem_cache *rmap_item_cache;
+static struct kmem_cache *mm_slot_cache;
+
+/* The number of nodes in the stable tree */
+static unsigned long ksm_kernel_pages_allocated;
+
+/* The number of page slots sharing those nodes */
+static unsigned long ksm_pages_shared;
+
+/* Limit on the number of unswappable pages used */
+static unsigned long ksm_max_kernel_pages;
+
+/* Number of pages ksmd should scan in one batch */
+static unsigned int ksm_thread_pages_to_scan;
+
+/* Milliseconds ksmd should sleep between batches */
+static unsigned int ksm_thread_sleep_millisecs;
+
+#define KSM_RUN_STOP   0
+#define KSM_RUN_MERGE  1
+#define KSM_RUN_UNMERGE        2
+static unsigned int ksm_run;
+
+static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
+static DEFINE_MUTEX(ksm_thread_mutex);
+static DEFINE_SPINLOCK(ksm_mmlist_lock);
+
+#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
+               sizeof(struct __struct), __alignof__(struct __struct),\
+               (__flags), NULL)
+
+static int __init ksm_slab_init(void)
+{
+       rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
+       if (!rmap_item_cache)
+               goto out;
+
+       mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
+       if (!mm_slot_cache)
+               goto out_free;
+
+       return 0;
+
+out_free:
+       kmem_cache_destroy(rmap_item_cache);
+out:
+       return -ENOMEM;
+}
+
+static void __init ksm_slab_free(void)
+{
+       kmem_cache_destroy(mm_slot_cache);
+       kmem_cache_destroy(rmap_item_cache);
+       mm_slot_cache = NULL;
+}
+
+static inline struct rmap_item *alloc_rmap_item(void)
+{
+       return kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
+}
+
+static inline void free_rmap_item(struct rmap_item *rmap_item)
+{
+       rmap_item->mm = NULL;   /* debug safety */
+       kmem_cache_free(rmap_item_cache, rmap_item);
+}
+
+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;
+}
+
+static void __init mm_slots_hash_free(void)
+{
+       kfree(mm_slots_hash);
+}
+
+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, link) {
+               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;
+       INIT_LIST_HEAD(&mm_slot->rmap_list);
+       hlist_add_head(&mm_slot->link, bucket);
+}
+
+static inline int in_stable_tree(struct rmap_item *rmap_item)
+{
+       return rmap_item->address & STABLE_FLAG;
+}
+
+/*
+ * We use break_ksm to break COW on a ksm page: it's a stripped down
+ *
+ *     if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
+ *             put_page(page);
+ *
+ * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
+ * in case the application has unmapped and remapped mm,addr meanwhile.
+ * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
+ * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
+ */
+static void break_ksm(struct vm_area_struct *vma, unsigned long addr)
+{
+       struct page *page;
+       int ret;
+
+       do {
+               cond_resched();
+               page = follow_page(vma, addr, FOLL_GET);
+               if (!page)
+                       break;
+               if (PageKsm(page))
+                       ret = handle_mm_fault(vma->vm_mm, vma, addr,
+                                                       FAULT_FLAG_WRITE);
+               else
+                       ret = VM_FAULT_WRITE;
+               put_page(page);
+       } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS)));
+
+       /* Which leaves us looping there if VM_FAULT_OOM: hmmm... */
+}
+
+static void __break_cow(struct mm_struct *mm, unsigned long addr)
+{
+       struct vm_area_struct *vma;
+
+       vma = find_vma(mm, addr);
+       if (!vma || vma->vm_start > addr)
+               return;
+       if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
+               return;
+       break_ksm(vma, addr);
+}
+
+static void break_cow(struct mm_struct *mm, unsigned long addr)
+{
+       down_read(&mm->mmap_sem);
+       __break_cow(mm, addr);
+       up_read(&mm->mmap_sem);
+}
+
+static struct page *get_mergeable_page(struct rmap_item *rmap_item)
+{
+       struct mm_struct *mm = rmap_item->mm;
+       unsigned long addr = rmap_item->address;
+       struct vm_area_struct *vma;
+       struct page *page;
+
+       down_read(&mm->mmap_sem);
+       vma = find_vma(mm, addr);
+       if (!vma || vma->vm_start > addr)
+               goto out;
+       if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
+               goto out;
+
+       page = follow_page(vma, addr, FOLL_GET);
+       if (!page)
+               goto out;
+       if (PageAnon(page)) {
+               flush_anon_page(vma, page, addr);
+               flush_dcache_page(page);
+       } else {
+               put_page(page);
+out:           page = NULL;
+       }
+       up_read(&mm->mmap_sem);
+       return page;
+}
+
+/*
+ * get_ksm_page: checks if the page at the virtual address in rmap_item
+ * is still PageKsm, in which case we can trust the content of the page,
+ * and it returns the gotten page; but NULL if the page has been zapped.
+ */
+static struct page *get_ksm_page(struct rmap_item *rmap_item)
+{
+       struct page *page;
+
+       page = get_mergeable_page(rmap_item);
+       if (page && !PageKsm(page)) {
+               put_page(page);
+               page = NULL;
+       }
+       return page;
+}
+
+/*
+ * Removing rmap_item from stable or unstable tree.
+ * This function will clean the information from the stable/unstable tree.
+ */
+static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
+{
+       if (in_stable_tree(rmap_item)) {
+               struct rmap_item *next_item = rmap_item->next;
+
+               if (rmap_item->address & NODE_FLAG) {
+                       if (next_item) {
+                               rb_replace_node(&rmap_item->node,
+                                               &next_item->node,
+                                               &root_stable_tree);
+                               next_item->address |= NODE_FLAG;
+                       } else {
+                               rb_erase(&rmap_item->node, &root_stable_tree);
+                               ksm_kernel_pages_allocated--;
+                       }
+               } else {
+                       struct rmap_item *prev_item = rmap_item->prev;
+
+                       BUG_ON(prev_item->next != rmap_item);
+                       prev_item->next = next_item;
+                       if (next_item) {
+                               BUG_ON(next_item->prev != rmap_item);
+                               next_item->prev = rmap_item->prev;
+                       }
+               }
+
+               rmap_item->next = NULL;
+               ksm_pages_shared--;
+
+       } else if (rmap_item->address & NODE_FLAG) {
+               unsigned char age;
+               /*
+                * ksm_thread can and must skip the rb_erase, because
+                * root_unstable_tree was already reset to RB_ROOT.
+                * But __ksm_exit has to be careful: do the rb_erase
+                * if it's interrupting a scan, and this rmap_item was
+                * inserted by this scan rather than left from before.
+                *
+                * Because of the case in which remove_mm_from_lists
+                * increments seqnr before removing rmaps, unstable_nr
+                * may even be 2 behind seqnr, but should never be
+                * further behind.  Yes, I did have trouble with this!
+                */
+               age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
+               BUG_ON(age > 2);
+               if (!age)
+                       rb_erase(&rmap_item->node, &root_unstable_tree);
+       }
+
+       rmap_item->address &= PAGE_MASK;
+
+       cond_resched();         /* we're called from many long loops */
+}
+
+static void remove_all_slot_rmap_items(struct mm_slot *mm_slot)
+{
+       struct rmap_item *rmap_item, *node;
+
+       list_for_each_entry_safe(rmap_item, node, &mm_slot->rmap_list, link) {
+               remove_rmap_item_from_tree(rmap_item);
+               list_del(&rmap_item->link);
+               free_rmap_item(rmap_item);
+       }
+}
+
+static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
+                                      struct list_head *cur)
+{
+       struct rmap_item *rmap_item;
+
+       while (cur != &mm_slot->rmap_list) {
+               rmap_item = list_entry(cur, struct rmap_item, link);
+               cur = cur->next;
+               remove_rmap_item_from_tree(rmap_item);
+               list_del(&rmap_item->link);
+               free_rmap_item(rmap_item);
+       }
+}
+
+/*
+ * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
+ * than check every pte of a given vma, the locking doesn't quite work for
+ * that - an rmap_item is assigned to the stable tree after inserting ksm
+ * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
+ * rmap_items from parent to child at fork time (so as not to waste time
+ * if exit comes before the next scan reaches it).
+ */
+static void unmerge_ksm_pages(struct vm_area_struct *vma,
+                             unsigned long start, unsigned long end)
+{
+       unsigned long addr;
+
+       for (addr = start; addr < end; addr += PAGE_SIZE)
+               break_ksm(vma, addr);
+}
+
+static void unmerge_and_remove_all_rmap_items(void)
+{
+       struct mm_slot *mm_slot;
+       struct mm_struct *mm;
+       struct vm_area_struct *vma;
+
+       list_for_each_entry(mm_slot, &ksm_mm_head.mm_list, mm_list) {
+               mm = mm_slot->mm;
+               down_read(&mm->mmap_sem);
+               for (vma = mm->mmap; vma; vma = vma->vm_next) {
+                       if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
+                               continue;
+                       unmerge_ksm_pages(vma, vma->vm_start, vma->vm_end);
+               }
+               remove_all_slot_rmap_items(mm_slot);
+               up_read(&mm->mmap_sem);
+       }
+
+       spin_lock(&ksm_mmlist_lock);
+       if (ksm_scan.mm_slot != &ksm_mm_head) {
+               ksm_scan.mm_slot = &ksm_mm_head;
+               ksm_scan.seqnr++;
+       }
+       spin_unlock(&ksm_mmlist_lock);
+}
+
+static void remove_mm_from_lists(struct mm_struct *mm)
+{
+       struct mm_slot *mm_slot;
+
+       spin_lock(&ksm_mmlist_lock);
+       mm_slot = get_mm_slot(mm);
+
+       /*
+        * This mm_slot is always at the scanning cursor when we're
+        * called from scan_get_next_rmap_item; but it's a special
+        * case when we're called from __ksm_exit.
+        */
+       if (ksm_scan.mm_slot == mm_slot) {
+               ksm_scan.mm_slot = list_entry(
+                       mm_slot->mm_list.next, struct mm_slot, mm_list);
+               ksm_scan.address = 0;
+               ksm_scan.rmap_item = list_entry(
+                       &ksm_scan.mm_slot->rmap_list, struct rmap_item, link);
+               if (ksm_scan.mm_slot == &ksm_mm_head)
+                       ksm_scan.seqnr++;
+       }
+
+       hlist_del(&mm_slot->link);
+       list_del(&mm_slot->mm_list);
+       spin_unlock(&ksm_mmlist_lock);
+
+       remove_all_slot_rmap_items(mm_slot);
+       free_mm_slot(mm_slot);
+       clear_bit(MMF_VM_MERGEABLE, &mm->flags);
+}
+
+static u32 calc_checksum(struct page *page)
+{
+       u32 checksum;
+       void *addr = kmap_atomic(page, KM_USER0);
+       checksum = jhash2(addr, PAGE_SIZE / 4, 17);
+       kunmap_atomic(addr, KM_USER0);
+       return checksum;
+}
+
+static int memcmp_pages(struct page *page1, struct page *page2)
+{
+       char *addr1, *addr2;
+       int ret;
+
+       addr1 = kmap_atomic(page1, KM_USER0);
+       addr2 = kmap_atomic(page2, KM_USER1);
+       ret = memcmp(addr1, addr2, PAGE_SIZE);
+       kunmap_atomic(addr2, KM_USER1);
+       kunmap_atomic(addr1, KM_USER0);
+       return ret;
+}
+
+static inline int pages_identical(struct page *page1, struct page *page2)
+{
+       return !memcmp_pages(page1, page2);
+}
+
+static int write_protect_page(struct vm_area_struct *vma, struct page *page,
+                             pte_t *orig_pte)
+{
+       struct mm_struct *mm = vma->vm_mm;
+       unsigned long addr;
+       pte_t *ptep;
+       spinlock_t *ptl;
+       int swapped;
+       int err = -EFAULT;
+
+       addr = page_address_in_vma(page, vma);
+       if (addr == -EFAULT)
+               goto out;
+
+       ptep = page_check_address(page, mm, addr, &ptl, 0);
+       if (!ptep)
+               goto out;
+
+       if (pte_write(*ptep)) {
+               pte_t entry;
+
+               swapped = PageSwapCache(page);
+               flush_cache_page(vma, addr, page_to_pfn(page));
+               /*
+                * Ok this is tricky, when get_user_pages_fast() run it doesnt
+                * take any lock, therefore the check that we are going to make
+                * with the pagecount against the mapcount is racey and
+                * O_DIRECT can happen right after the check.
+                * So we clear the pte and flush the tlb before the check
+                * this assure us that no O_DIRECT can happen after the check
+                * or in the middle of the check.
+                */
+               entry = ptep_clear_flush(vma, addr, ptep);
+               /*
+                * Check that no O_DIRECT or similar I/O is in progress on the
+                * page
+                */
+               if ((page_mapcount(page) + 2 + swapped) != page_count(page)) {
+                       set_pte_at_notify(mm, addr, ptep, entry);
+                       goto out_unlock;
+               }
+               entry = pte_wrprotect(entry);
+               set_pte_at_notify(mm, addr, ptep, entry);
+       }
+       *orig_pte = *ptep;
+       err = 0;
+
+out_unlock:
+       pte_unmap_unlock(ptep, ptl);
+out:
+       return err;
+}
+
+/**
+ * replace_page - replace page in vma by new ksm page
+ * @vma:      vma that holds the pte pointing to oldpage
+ * @oldpage:  the page we are replacing by newpage
+ * @newpage:  the ksm page we replace oldpage by
+ * @orig_pte: the original value of the pte
+ *
+ * Returns 0 on success, -EFAULT on failure.
+ */
+static int replace_page(struct vm_area_struct *vma, struct page *oldpage,
+                       struct page *newpage, pte_t orig_pte)
+{
+       struct mm_struct *mm = vma->vm_mm;
+       pgd_t *pgd;
+       pud_t *pud;
+       pmd_t *pmd;
+       pte_t *ptep;
+       spinlock_t *ptl;
+       unsigned long addr;
+       pgprot_t prot;
+       int err = -EFAULT;
+
+       prot = vm_get_page_prot(vma->vm_flags & ~VM_WRITE);
+
+       addr = page_address_in_vma(oldpage, vma);
+       if (addr == -EFAULT)
+               goto out;
+
+       pgd = pgd_offset(mm, addr);
+       if (!pgd_present(*pgd))
+               goto out;
+
+       pud = pud_offset(pgd, addr);
+       if (!pud_present(*pud))
+               goto out;
+
+       pmd = pmd_offset(pud, addr);
+       if (!pmd_present(*pmd))
+               goto out;
+
+       ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
+       if (!pte_same(*ptep, orig_pte)) {
+               pte_unmap_unlock(ptep, ptl);
+               goto out;
+       }
+
+       get_page(newpage);
+       page_add_ksm_rmap(newpage);
+
+       flush_cache_page(vma, addr, pte_pfn(*ptep));
+       ptep_clear_flush(vma, addr, ptep);
+       set_pte_at_notify(mm, addr, ptep, mk_pte(newpage, prot));
+
+       page_remove_rmap(oldpage);
+       put_page(oldpage);
+
+       pte_unmap_unlock(ptep, ptl);
+       err = 0;
+out:
+       return err;
+}
+
+/*
+ * try_to_merge_one_page - take two pages and merge them into one
+ * @vma: the vma that hold the pte pointing into oldpage
+ * @oldpage: the page that we want to replace with newpage
+ * @newpage: the page that we want to map instead of oldpage
+ *
+ * Note:
+ * oldpage should be a PageAnon page, while newpage should be a PageKsm page,
+ * or a newly allocated kernel page which page_add_ksm_rmap will make PageKsm.
+ *
+ * This function returns 0 if the pages were merged, -EFAULT otherwise.
+ */
+static int try_to_merge_one_page(struct vm_area_struct *vma,
+                                struct page *oldpage,
+                                struct page *newpage)
+{
+       pte_t orig_pte = __pte(0);
+       int err = -EFAULT;
+
+       if (!(vma->vm_flags & VM_MERGEABLE))
+               goto out;
+
+       if (!PageAnon(oldpage))
+               goto out;
+
+       get_page(newpage);
+       get_page(oldpage);
+
+       /*
+        * We need the page lock to read a stable PageSwapCache in
+        * write_protect_page().  We use trylock_page() instead of
+        * lock_page() because we don't want to wait here - we
+        * prefer to continue scanning and merging different pages,
+        * then come back to this page when it is unlocked.
+        */
+       if (!trylock_page(oldpage))
+               goto out_putpage;
+       /*
+        * If this anonymous page is mapped only here, its pte may need
+        * to be write-protected.  If it's mapped elsewhere, all of its
+        * ptes are necessarily already write-protected.  But in either
+        * case, we need to lock and check page_count is not raised.
+        */
+       if (write_protect_page(vma, oldpage, &orig_pte)) {
+               unlock_page(oldpage);
+               goto out_putpage;
+       }
+       unlock_page(oldpage);
+
+       if (pages_identical(oldpage, newpage))
+               err = replace_page(vma, oldpage, newpage, orig_pte);
+
+out_putpage:
+       put_page(oldpage);
+       put_page(newpage);
+out:
+       return err;
+}
+
+/*
+ * try_to_merge_two_pages - take two identical pages and prepare them
+ * to be merged into one page.
+ *
+ * This function returns 0 if we successfully mapped two identical pages
+ * into one page, -EFAULT otherwise.
+ *
+ * Note that this function allocates a new kernel page: if one of the pages
+ * is already a ksm page, try_to_merge_with_ksm_page should be used.
+ */
+static int try_to_merge_two_pages(struct mm_struct *mm1, unsigned long addr1,
+                                 struct page *page1, struct mm_struct *mm2,
+                                 unsigned long addr2, struct page *page2)
+{
+       struct vm_area_struct *vma;
+       struct page *kpage;
+       int err = -EFAULT;
+
+       /*
+        * The number of nodes in the stable tree
+        * is the number of kernel pages that we hold.
+        */
+       if (ksm_max_kernel_pages &&
+           ksm_max_kernel_pages <= ksm_kernel_pages_allocated)
+               return err;
+
+       kpage = alloc_page(GFP_HIGHUSER);
+       if (!kpage)
+               return err;
+
+       down_read(&mm1->mmap_sem);
+       vma = find_vma(mm1, addr1);
+       if (!vma || vma->vm_start > addr1) {
+               put_page(kpage);
+               up_read(&mm1->mmap_sem);
+               return err;
+       }
+
+       copy_user_highpage(kpage, page1, addr1, vma);
+       err = try_to_merge_one_page(vma, page1, kpage);
+       up_read(&mm1->mmap_sem);
+
+       if (!err) {
+               down_read(&mm2->mmap_sem);
+               vma = find_vma(mm2, addr2);
+               if (!vma || vma->vm_start > addr2) {
+                       put_page(kpage);
+                       up_read(&mm2->mmap_sem);
+                       break_cow(mm1, addr1);
+                       return -EFAULT;
+               }
+
+               err = try_to_merge_one_page(vma, page2, kpage);
+               up_read(&mm2->mmap_sem);
+
+               /*
+                * If the second try_to_merge_one_page failed, we have a
+                * ksm page with just one pte pointing to it, so break it.
+                */
+               if (err)
+                       break_cow(mm1, addr1);
+               else
+                       ksm_pages_shared += 2;
+       }
+
+       put_page(kpage);
+       return err;
+}
+
+/*
+ * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
+ * but no new kernel page is allocated: kpage must already be a ksm page.
+ */
+static int try_to_merge_with_ksm_page(struct mm_struct *mm1,
+                                     unsigned long addr1,
+                                     struct page *page1,
+                                     struct page *kpage)
+{
+       struct vm_area_struct *vma;
+       int err = -EFAULT;
+
+       down_read(&mm1->mmap_sem);
+       vma = find_vma(mm1, addr1);
+       if (!vma || vma->vm_start > addr1) {
+               up_read(&mm1->mmap_sem);
+               return err;
+       }
+
+       err = try_to_merge_one_page(vma, page1, kpage);
+       up_read(&mm1->mmap_sem);
+
+       if (!err)
+               ksm_pages_shared++;
+
+       return err;
+}
+
+/*
+ * stable_tree_search - search page inside the stable tree
+ * @page: the page that we are searching identical pages to.
+ * @page2: pointer into identical page that we are holding inside the stable
+ *        tree that we have found.
+ * @rmap_item: the reverse mapping item
+ *
+ * This function checks if there is a page inside the stable tree
+ * with identical content to the page that we are scanning right now.
+ *
+ * This function return rmap_item pointer to the identical item if found,
+ * NULL otherwise.
+ */
+static struct rmap_item *stable_tree_search(struct page *page,
+                                           struct page **page2,
+                                           struct rmap_item *rmap_item)
+{
+       struct rb_node *node = root_stable_tree.rb_node;
+
+       while (node) {
+               struct rmap_item *tree_rmap_item, *next_rmap_item;
+               int ret;
+
+               tree_rmap_item = rb_entry(node, struct rmap_item, node);
+               while (tree_rmap_item) {
+                       BUG_ON(!in_stable_tree(tree_rmap_item));
+                       cond_resched();
+                       page2[0] = get_ksm_page(tree_rmap_item);
+                       if (page2[0])
+                               break;
+                       next_rmap_item = tree_rmap_item->next;
+                       remove_rmap_item_from_tree(tree_rmap_item);
+                       tree_rmap_item = next_rmap_item;
+               }
+               if (!tree_rmap_item)
+                       return NULL;
+
+               ret = memcmp_pages(page, page2[0]);
+
+               if (ret < 0) {
+                       put_page(page2[0]);
+                       node = node->rb_left;
+               } else if (ret > 0) {
+                       put_page(page2[0]);
+                       node = node->rb_right;
+               } else {
+                       return tree_rmap_item;
+               }
+       }
+
+       return NULL;
+}
+
+/*
+ * stable_tree_insert - insert rmap_item pointing to new ksm page
+ * into the stable tree.
+ *
+ * @page: the page that we are searching identical page to inside the stable
+ *       tree.
+ * @rmap_item: pointer to the reverse mapping item.
+ *
+ * This function returns rmap_item if success, NULL otherwise.
+ */
+static struct rmap_item *stable_tree_insert(struct page *page,
+                                           struct rmap_item *rmap_item)
+{
+       struct rb_node **new = &root_stable_tree.rb_node;
+       struct rb_node *parent = NULL;
+
+       while (*new) {
+               struct rmap_item *tree_rmap_item, *next_rmap_item;
+               struct page *tree_page;
+               int ret;
+
+               tree_rmap_item = rb_entry(*new, struct rmap_item, node);
+               while (tree_rmap_item) {
+                       BUG_ON(!in_stable_tree(tree_rmap_item));
+                       cond_resched();
+                       tree_page = get_ksm_page(tree_rmap_item);
+                       if (tree_page)
+                               break;
+                       next_rmap_item = tree_rmap_item->next;
+                       remove_rmap_item_from_tree(tree_rmap_item);
+                       tree_rmap_item = next_rmap_item;
+               }
+               if (!tree_rmap_item)
+                       return NULL;
+
+               ret = memcmp_pages(page, tree_page);
+               put_page(tree_page);
+
+               parent = *new;
+               if (ret < 0)
+                       new = &parent->rb_left;
+               else if (ret > 0)
+                       new = &parent->rb_right;
+               else {
+                       /*
+                        * It is not a bug that stable_tree_search() didn't
+                        * find this node: because at that time our page was
+                        * not yet write-protected, so may have changed since.
+                        */
+                       return NULL;
+               }
+       }
+
+       ksm_kernel_pages_allocated++;
+
+       rmap_item->address |= NODE_FLAG | STABLE_FLAG;
+       rmap_item->next = NULL;
+       rb_link_node(&rmap_item->node, parent, new);
+       rb_insert_color(&rmap_item->node, &root_stable_tree);
+
+       return rmap_item;
+}
+
+/*
+ * unstable_tree_search_insert - search and insert items into the unstable tree.
+ *
+ * @page: the page that we are going to search for identical page or to insert
+ *       into the unstable tree
+ * @page2: pointer into identical page that was found inside the unstable tree
+ * @rmap_item: the reverse mapping item of page
+ *
+ * This function searches for a page in the unstable tree identical to the
+ * page currently being scanned; and if no identical page is found in the
+ * tree, we insert rmap_item as a new object into the unstable tree.
+ *
+ * This function returns pointer to rmap_item found to be identical
+ * to the currently scanned page, NULL otherwise.
+ *
+ * This function does both searching and inserting, because they share
+ * the same walking algorithm in an rbtree.
+ */
+static struct rmap_item *unstable_tree_search_insert(struct page *page,
+                                               struct page **page2,
+                                               struct rmap_item *rmap_item)
+{
+       struct rb_node **new = &root_unstable_tree.rb_node;
+       struct rb_node *parent = NULL;
+
+       while (*new) {
+               struct rmap_item *tree_rmap_item;
+               int ret;
+
+               tree_rmap_item = rb_entry(*new, struct rmap_item, node);
+               page2[0] = get_mergeable_page(tree_rmap_item);
+               if (!page2[0])
+                       return NULL;
+
+               /*
+                * Don't substitute an unswappable ksm page
+                * just for one good swappable forked page.
+                */
+               if (page == page2[0]) {
+                       put_page(page2[0]);
+                       return NULL;
+               }
+
+               ret = memcmp_pages(page, page2[0]);
+
+               parent = *new;
+               if (ret < 0) {
+                       put_page(page2[0]);
+                       new = &parent->rb_left;
+               } else if (ret > 0) {
+                       put_page(page2[0]);
+                       new = &parent->rb_right;
+               } else {
+                       return tree_rmap_item;
+               }
+       }
+
+       rmap_item->address |= NODE_FLAG;
+       rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
+       rb_link_node(&rmap_item->node, parent, new);
+       rb_insert_color(&rmap_item->node, &root_unstable_tree);
+
+       return NULL;
+}
+
+/*
+ * stable_tree_append - add another rmap_item to the linked list of
+ * rmap_items hanging off a given node of the stable tree, all sharing
+ * the same ksm page.
+ */
+static void stable_tree_append(struct rmap_item *rmap_item,
+                              struct rmap_item *tree_rmap_item)
+{
+       rmap_item->next = tree_rmap_item->next;
+       rmap_item->prev = tree_rmap_item;
+
+       if (tree_rmap_item->next)
+               tree_rmap_item->next->prev = rmap_item;
+
+       tree_rmap_item->next = rmap_item;
+       rmap_item->address |= STABLE_FLAG;
+}
+
+/*
+ * cmp_and_merge_page - take a page computes its hash value and check if there
+ * is similar hash value to different page,
+ * in case we find that there is similar hash to different page we call to
+ * try_to_merge_two_pages().
+ *
+ * @page: the page that we are searching identical page to.
+ * @rmap_item: the reverse mapping into the virtual address of this page
+ */
+static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
+{
+       struct page *page2[1];
+       struct rmap_item *tree_rmap_item;
+       unsigned int checksum;
+       int err;
+
+       if (in_stable_tree(rmap_item))
+               remove_rmap_item_from_tree(rmap_item);
+
+       /* We first start with searching the page inside the stable tree */
+       tree_rmap_item = stable_tree_search(page, page2, rmap_item);
+       if (tree_rmap_item) {
+               if (page == page2[0]) {                 /* forked */
+                       ksm_pages_shared++;
+                       err = 0;
+               } else
+                       err = try_to_merge_with_ksm_page(rmap_item->mm,
+                                                        rmap_item->address,
+                                                        page, page2[0]);
+               put_page(page2[0]);
+
+               if (!err) {
+                       /*
+                        * The page was successfully merged:
+                        * add its rmap_item to the stable tree.
+                        */
+                       stable_tree_append(rmap_item, tree_rmap_item);
+               }
+               return;
+       }
+
+       /*
+        * A ksm page might have got here by fork, but its other
+        * references have already been removed from the stable tree.
+        */
+       if (PageKsm(page))
+               break_cow(rmap_item->mm, rmap_item->address);
+
+       /*
+        * In case the hash value of the page was changed from the last time we
+        * have calculated it, this page to be changed frequely, therefore we
+        * don't want to insert it to the unstable tree, and we don't want to
+        * waste our time to search if there is something identical to it there.
+        */
+       checksum = calc_checksum(page);
+       if (rmap_item->oldchecksum != checksum) {
+               rmap_item->oldchecksum = checksum;
+               return;
+       }
+
+       tree_rmap_item = unstable_tree_search_insert(page, page2, rmap_item);
+       if (tree_rmap_item) {
+               err = try_to_merge_two_pages(rmap_item->mm,
+                                            rmap_item->address, page,
+                                            tree_rmap_item->mm,
+                                            tree_rmap_item->address, page2[0]);
+               /*
+                * As soon as we merge this page, we want to remove the
+                * rmap_item of the page we have merged with from the unstable
+                * tree, and insert it instead as new node in the stable tree.
+                */
+               if (!err) {
+                       rb_erase(&tree_rmap_item->node, &root_unstable_tree);
+                       tree_rmap_item->address &= ~NODE_FLAG;
+                       /*
+                        * If we fail to insert the page into the stable tree,
+                        * we will have 2 virtual addresses that are pointing
+                        * to a ksm page left outside the stable tree,
+                        * in which case we need to break_cow on both.
+                        */
+                       if (stable_tree_insert(page2[0], tree_rmap_item))
+                               stable_tree_append(rmap_item, tree_rmap_item);
+                       else {
+                               break_cow(tree_rmap_item->mm,
+                                               tree_rmap_item->address);
+                               break_cow(rmap_item->mm, rmap_item->address);
+                               ksm_pages_shared -= 2;
+                       }
+               }
+
+               put_page(page2[0]);
+       }
+}
+
+static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
+                                           struct list_head *cur,
+                                           unsigned long addr)
+{
+       struct rmap_item *rmap_item;
+
+       while (cur != &mm_slot->rmap_list) {
+               rmap_item = list_entry(cur, struct rmap_item, link);
+               if ((rmap_item->address & PAGE_MASK) == addr) {
+                       if (!in_stable_tree(rmap_item))
+                               remove_rmap_item_from_tree(rmap_item);
+                       return rmap_item;
+               }
+               if (rmap_item->address > addr)
+                       break;
+               cur = cur->next;
+               remove_rmap_item_from_tree(rmap_item);
+               list_del(&rmap_item->link);
+               free_rmap_item(rmap_item);
+       }
+
+       rmap_item = alloc_rmap_item();
+       if (rmap_item) {
+               /* It has already been zeroed */
+               rmap_item->mm = mm_slot->mm;
+               rmap_item->address = addr;
+               list_add_tail(&rmap_item->link, cur);
+       }
+       return rmap_item;
+}
+
+static struct rmap_item *scan_get_next_rmap_item(struct page **page)
+{
+       struct mm_struct *mm;
+       struct mm_slot *slot;
+       struct vm_area_struct *vma;
+       struct rmap_item *rmap_item;
+
+       if (list_empty(&ksm_mm_head.mm_list))
+               return NULL;
+
+       slot = ksm_scan.mm_slot;
+       if (slot == &ksm_mm_head) {
+               root_unstable_tree = RB_ROOT;
+
+               spin_lock(&ksm_mmlist_lock);
+               slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
+               ksm_scan.mm_slot = slot;
+               spin_unlock(&ksm_mmlist_lock);
+next_mm:
+               ksm_scan.address = 0;
+               ksm_scan.rmap_item = list_entry(&slot->rmap_list,
+                                               struct rmap_item, link);
+       }
+
+       mm = slot->mm;
+       down_read(&mm->mmap_sem);
+       for (vma = find_vma(mm, ksm_scan.address); vma; vma = vma->vm_next) {
+               if (!(vma->vm_flags & VM_MERGEABLE))
+                       continue;
+               if (ksm_scan.address < vma->vm_start)
+                       ksm_scan.address = vma->vm_start;
+               if (!vma->anon_vma)
+                       ksm_scan.address = vma->vm_end;
+
+               while (ksm_scan.address < vma->vm_end) {
+                       *page = follow_page(vma, ksm_scan.address, FOLL_GET);
+                       if (*page && PageAnon(*page)) {
+                               flush_anon_page(vma, *page, ksm_scan.address);
+                               flush_dcache_page(*page);
+                               rmap_item = get_next_rmap_item(slot,
+                                       ksm_scan.rmap_item->link.next,
+                                       ksm_scan.address);
+                               if (rmap_item) {
+                                       ksm_scan.rmap_item = rmap_item;
+                                       ksm_scan.address += PAGE_SIZE;
+                               } else
+                                       put_page(*page);
+                               up_read(&mm->mmap_sem);
+                               return rmap_item;
+                       }
+                       if (*page)
+                               put_page(*page);
+                       ksm_scan.address += PAGE_SIZE;
+                       cond_resched();
+               }
+       }
+
+       if (!ksm_scan.address) {
+               /*
+                * We've completed a full scan of all vmas, holding mmap_sem
+                * throughout, and found no VM_MERGEABLE: so do the same as
+                * __ksm_exit does to remove this mm from all our lists now.
+                */
+               remove_mm_from_lists(mm);
+               up_read(&mm->mmap_sem);
+               slot = ksm_scan.mm_slot;
+               if (slot != &ksm_mm_head)
+                       goto next_mm;
+               return NULL;
+       }
+
+       /*
+        * Nuke all the rmap_items that are above this current rmap:
+        * because there were no VM_MERGEABLE vmas with such addresses.
+        */
+       remove_trailing_rmap_items(slot, ksm_scan.rmap_item->link.next);
+       up_read(&mm->mmap_sem);
+
+       spin_lock(&ksm_mmlist_lock);
+       slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
+       ksm_scan.mm_slot = slot;
+       spin_unlock(&ksm_mmlist_lock);
+
+       /* Repeat until we've completed scanning the whole list */
+       if (slot != &ksm_mm_head)
+               goto next_mm;
+
+       /*
+        * Bump seqnr here rather than at top, so that __ksm_exit
+        * can skip rb_erase on unstable tree until we run again.
+        */
+       ksm_scan.seqnr++;
+       return NULL;
+}
+
+/**
+ * ksm_do_scan  - the ksm scanner main worker function.
+ * @scan_npages - number of pages we want to scan before we return.
+ */
+static void ksm_do_scan(unsigned int scan_npages)
+{
+       struct rmap_item *rmap_item;
+       struct page *page;
+
+       while (scan_npages--) {
+               cond_resched();
+               rmap_item = scan_get_next_rmap_item(&page);
+               if (!rmap_item)
+                       return;
+               if (!PageKsm(page) || !in_stable_tree(rmap_item))
+                       cmp_and_merge_page(page, rmap_item);
+               put_page(page);
+       }
+}
+
+static int ksm_scan_thread(void *nothing)
+{
+       set_user_nice(current, 0);
+
+       while (!kthread_should_stop()) {
+               if (ksm_run & KSM_RUN_MERGE) {
+                       mutex_lock(&ksm_thread_mutex);
+                       ksm_do_scan(ksm_thread_pages_to_scan);
+                       mutex_unlock(&ksm_thread_mutex);
+                       schedule_timeout_interruptible(
+                               msecs_to_jiffies(ksm_thread_sleep_millisecs));
+               } else {
+                       wait_event_interruptible(ksm_thread_wait,
+                                       (ksm_run & KSM_RUN_MERGE) ||
+                                       kthread_should_stop());
+               }
+       }
+       return 0;
+}
+
 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
                unsigned long end, int advice, unsigned long *vm_flags)
 {
@@ -33,7 +1307,8 @@ int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
                if (!(*vm_flags & VM_MERGEABLE))
                        return 0;               /* just ignore the advice */
 
-               /* Unmerge any merged pages here */
+               if (vma->anon_vma)
+                       unmerge_ksm_pages(vma, start, end);
 
                *vm_flags &= ~VM_MERGEABLE;
                break;
@@ -44,13 +1319,217 @@ int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
 
 int __ksm_enter(struct mm_struct *mm)
 {
-       /* Allocate a structure to track mm and link it into KSM's list */
+       struct mm_slot *mm_slot = alloc_mm_slot();
+       if (!mm_slot)
+               return -ENOMEM;
+
+       spin_lock(&ksm_mmlist_lock);
+       insert_to_mm_slots_hash(mm, mm_slot);
+       /*
+        * Insert just behind the scanning cursor, to let the area settle
+        * down a little; when fork is followed by immediate exec, we don't
+        * want ksmd to waste time setting up and tearing down an rmap_list.
+        */
+       list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
+       spin_unlock(&ksm_mmlist_lock);
+
        set_bit(MMF_VM_MERGEABLE, &mm->flags);
        return 0;
 }
 
 void __ksm_exit(struct mm_struct *mm)
 {
-       /* Unlink and free all KSM's structures which track this mm */
-       clear_bit(MMF_VM_MERGEABLE, &mm->flags);
+       /*
+        * This process is exiting: doesn't hold and doesn't need mmap_sem;
+        * but we do need to exclude ksmd and other exiters while we modify
+        * the various lists and trees.
+        */
+       mutex_lock(&ksm_thread_mutex);
+       remove_mm_from_lists(mm);
+       mutex_unlock(&ksm_thread_mutex);
+}
+
+#define KSM_ATTR_RO(_name) \
+       static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
+#define KSM_ATTR(_name) \
+       static struct kobj_attribute _name##_attr = \
+               __ATTR(_name, 0644, _name##_show, _name##_store)
+
+static ssize_t sleep_millisecs_show(struct kobject *kobj,
+                                   struct kobj_attribute *attr, char *buf)
+{
+       return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
+}
+
+static ssize_t 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;
+
+       ksm_thread_sleep_millisecs = msecs;
+
+       return count;
+}
+KSM_ATTR(sleep_millisecs);
+
+static ssize_t pages_to_scan_show(struct kobject *kobj,
+                                 struct kobj_attribute *attr, char *buf)
+{
+       return sprintf(buf, "%u\n", ksm_thread_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 nr_pages;
+
+       err = strict_strtoul(buf, 10, &nr_pages);
+       if (err || nr_pages > UINT_MAX)
+               return -EINVAL;
+
+       ksm_thread_pages_to_scan = nr_pages;
+
+       return count;
+}
+KSM_ATTR(pages_to_scan);
+
+static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
+                       char *buf)
+{
+       return sprintf(buf, "%u\n", ksm_run);
+}
+
+static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
+                        const char *buf, size_t count)
+{
+       int err;
+       unsigned long flags;
+
+       err = strict_strtoul(buf, 10, &flags);
+       if (err || flags > UINT_MAX)
+               return -EINVAL;
+       if (flags > KSM_RUN_UNMERGE)
+               return -EINVAL;
+
+       /*
+        * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
+        * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
+        * breaking COW to free the kernel_pages_allocated (but leaves
+        * mm_slots on the list for when ksmd may be set running again).
+        */
+
+       mutex_lock(&ksm_thread_mutex);
+       if (ksm_run != flags) {
+               ksm_run = flags;
+               if (flags & KSM_RUN_UNMERGE)
+                       unmerge_and_remove_all_rmap_items();
+       }
+       mutex_unlock(&ksm_thread_mutex);
+
+       if (flags & KSM_RUN_MERGE)
+               wake_up_interruptible(&ksm_thread_wait);
+
+       return count;
+}
+KSM_ATTR(run);
+
+static ssize_t pages_shared_show(struct kobject *kobj,
+                                struct kobj_attribute *attr, char *buf)
+{
+       return sprintf(buf, "%lu\n",
+                       ksm_pages_shared - ksm_kernel_pages_allocated);
+}
+KSM_ATTR_RO(pages_shared);
+
+static ssize_t kernel_pages_allocated_show(struct kobject *kobj,
+                                          struct kobj_attribute *attr,
+                                          char *buf)
+{
+       return sprintf(buf, "%lu\n", ksm_kernel_pages_allocated);
+}
+KSM_ATTR_RO(kernel_pages_allocated);
+
+static ssize_t max_kernel_pages_store(struct kobject *kobj,
+                                     struct kobj_attribute *attr,
+                                     const char *buf, size_t count)
+{
+       int err;
+       unsigned long nr_pages;
+
+       err = strict_strtoul(buf, 10, &nr_pages);
+       if (err)
+               return -EINVAL;
+
+       ksm_max_kernel_pages = nr_pages;
+
+       return count;
+}
+
+static ssize_t max_kernel_pages_show(struct kobject *kobj,
+                                    struct kobj_attribute *attr, char *buf)
+{
+       return sprintf(buf, "%lu\n", ksm_max_kernel_pages);
+}
+KSM_ATTR(max_kernel_pages);
+
+static struct attribute *ksm_attrs[] = {
+       &sleep_millisecs_attr.attr,
+       &pages_to_scan_attr.attr,
+       &run_attr.attr,
+       &pages_shared_attr.attr,
+       &kernel_pages_allocated_attr.attr,
+       &max_kernel_pages_attr.attr,
+       NULL,
+};
+
+static struct attribute_group ksm_attr_group = {
+       .attrs = ksm_attrs,
+       .name = "ksm",
+};
+
+static int __init ksm_init(void)
+{
+       struct task_struct *ksm_thread;
+       int err;
+
+       err = ksm_slab_init();
+       if (err)
+               goto out;
+
+       err = mm_slots_hash_init();
+       if (err)
+               goto out_free1;
+
+       ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
+       if (IS_ERR(ksm_thread)) {
+               printk(KERN_ERR "ksm: creating kthread failed\n");
+               err = PTR_ERR(ksm_thread);
+               goto out_free2;
+       }
+
+       err = sysfs_create_group(mm_kobj, &ksm_attr_group);
+       if (err) {
+               printk(KERN_ERR "ksm: register sysfs failed\n");
+               goto out_free3;
+       }
+
+       return 0;
+
+out_free3:
+       kthread_stop(ksm_thread);
+out_free2:
+       mm_slots_hash_free();
+out_free1:
+       ksm_slab_free();
+out:
+       return err;
 }
+module_init(ksm_init)