memcg: charge swapcache to proper memcg
[linux-2.6.git] / mm / memcontrol.c
index ab2ecbb..81b0ae8 100644 (file)
@@ -27,6 +27,7 @@
 #include <linux/backing-dev.h>
 #include <linux/bit_spinlock.h>
 #include <linux/rcupdate.h>
+#include <linux/limits.h>
 #include <linux/mutex.h>
 #include <linux/slab.h>
 #include <linux/swap.h>
@@ -51,6 +52,7 @@ static int really_do_swap_account __initdata = 1; /* for remember boot option*/
 #define do_swap_account                (0)
 #endif
 
+static DEFINE_MUTEX(memcg_tasklist);   /* can be hold under cgroup_mutex */
 
 /*
  * Statistics for memory cgroup.
@@ -94,6 +96,15 @@ static s64 mem_cgroup_read_stat(struct mem_cgroup_stat *stat,
        return ret;
 }
 
+static s64 mem_cgroup_local_usage(struct mem_cgroup_stat *stat)
+{
+       s64 ret;
+
+       ret = mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_CACHE);
+       ret += mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_RSS);
+       return ret;
+}
+
 /*
  * per-zone information in memory controller.
  */
@@ -153,22 +164,18 @@ struct mem_cgroup {
 
        /*
         * While reclaiming in a hiearchy, we cache the last child we
-        * reclaimed from. Protected by cgroup_lock()
+        * reclaimed from.
         */
-       struct mem_cgroup *last_scanned_child;
+       int last_scanned_child;
        /*
         * Should the accounting and control be hierarchical, per subtree?
         */
        bool use_hierarchy;
        unsigned long   last_oom_jiffies;
-       int             obsolete;
        atomic_t        refcnt;
 
        unsigned int    swappiness;
 
-
-       unsigned int inactive_ratio;
-
        /*
         * statistics. This must be placed at the end of memcg.
         */
@@ -205,6 +212,7 @@ pcg_default_flags[NR_CHARGE_TYPE] = {
 
 static void mem_cgroup_get(struct mem_cgroup *mem);
 static void mem_cgroup_put(struct mem_cgroup *mem);
+static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
 
 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
                                         struct page_cgroup *pc,
@@ -249,7 +257,7 @@ page_cgroup_zoneinfo(struct page_cgroup *pc)
        return mem_cgroup_zoneinfo(mem, nid, zid);
 }
 
-static unsigned long mem_cgroup_get_all_zonestat(struct mem_cgroup *mem,
+static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
                                        enum lru_list idx)
 {
        int nid, zid;
@@ -285,6 +293,70 @@ struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
                                struct mem_cgroup, css);
 }
 
+static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
+{
+       struct mem_cgroup *mem = NULL;
+
+       if (!mm)
+               return NULL;
+       /*
+        * Because we have no locks, mm->owner's may be being moved to other
+        * cgroup. We use css_tryget() here even if this looks
+        * pessimistic (rather than adding locks here).
+        */
+       rcu_read_lock();
+       do {
+               mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
+               if (unlikely(!mem))
+                       break;
+       } while (!css_tryget(&mem->css));
+       rcu_read_unlock();
+       return mem;
+}
+
+static bool mem_cgroup_is_obsolete(struct mem_cgroup *mem)
+{
+       if (!mem)
+               return true;
+       return css_is_removed(&mem->css);
+}
+
+
+/*
+ * Call callback function against all cgroup under hierarchy tree.
+ */
+static int mem_cgroup_walk_tree(struct mem_cgroup *root, void *data,
+                         int (*func)(struct mem_cgroup *, void *))
+{
+       int found, ret, nextid;
+       struct cgroup_subsys_state *css;
+       struct mem_cgroup *mem;
+
+       if (!root->use_hierarchy)
+               return (*func)(root, data);
+
+       nextid = 1;
+       do {
+               ret = 0;
+               mem = NULL;
+
+               rcu_read_lock();
+               css = css_get_next(&mem_cgroup_subsys, nextid, &root->css,
+                                  &found);
+               if (css && css_tryget(css))
+                       mem = container_of(css, struct mem_cgroup, css);
+               rcu_read_unlock();
+
+               if (mem) {
+                       ret = (*func)(mem, data);
+                       css_put(&mem->css);
+               }
+               nextid = found + 1;
+       } while (!ret && css);
+
+       return ret;
+}
+
 /*
  * Following LRU functions are allowed to be used without PCG_LOCK.
  * Operations are called by routine of global LRU independently from memcg.
@@ -309,8 +381,12 @@ void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
                return;
        pc = lookup_page_cgroup(page);
        /* can happen while we handle swapcache. */
-       if (list_empty(&pc->lru))
+       if (list_empty(&pc->lru) || !pc->mem_cgroup)
                return;
+       /*
+        * We don't check PCG_USED bit. It's cleared when the "page" is finally
+        * removed from global LRU.
+        */
        mz = page_cgroup_zoneinfo(pc);
        mem = pc->mem_cgroup;
        MEM_CGROUP_ZSTAT(mz, lru) -= 1;
@@ -332,6 +408,10 @@ void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
                return;
 
        pc = lookup_page_cgroup(page);
+       /*
+        * Used bit is set without atomic ops but after smp_wmb().
+        * For making pc->mem_cgroup visible, insert smp_rmb() here.
+        */
        smp_rmb();
        /* unused page is not rotated. */
        if (!PageCgroupUsed(pc))
@@ -348,7 +428,10 @@ void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
        if (mem_cgroup_disabled())
                return;
        pc = lookup_page_cgroup(page);
-       /* barrier to sync with "charge" */
+       /*
+        * Used bit is set without atomic ops but after smp_wmb().
+        * For making pc->mem_cgroup visible, insert smp_rmb() here.
+        */
        smp_rmb();
        if (!PageCgroupUsed(pc))
                return;
@@ -357,16 +440,44 @@ void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
        MEM_CGROUP_ZSTAT(mz, lru) += 1;
        list_add(&pc->lru, &mz->lists[lru]);
 }
+
 /*
- * To add swapcache into LRU. Be careful to all this function.
- * zone->lru_lock shouldn't be held and irq must not be disabled.
+ * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
+ * lru because the page may.be reused after it's fully uncharged (because of
+ * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
+ * it again. This function is only used to charge SwapCache. It's done under
+ * lock_page and expected that zone->lru_lock is never held.
  */
-static void mem_cgroup_lru_fixup(struct page *page)
+static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
+{
+       unsigned long flags;
+       struct zone *zone = page_zone(page);
+       struct page_cgroup *pc = lookup_page_cgroup(page);
+
+       spin_lock_irqsave(&zone->lru_lock, flags);
+       /*
+        * Forget old LRU when this page_cgroup is *not* used. This Used bit
+        * is guarded by lock_page() because the page is SwapCache.
+        */
+       if (!PageCgroupUsed(pc))
+               mem_cgroup_del_lru_list(page, page_lru(page));
+       spin_unlock_irqrestore(&zone->lru_lock, flags);
+}
+
+static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
 {
-       if (!isolate_lru_page(page))
-               putback_lru_page(page);
+       unsigned long flags;
+       struct zone *zone = page_zone(page);
+       struct page_cgroup *pc = lookup_page_cgroup(page);
+
+       spin_lock_irqsave(&zone->lru_lock, flags);
+       /* link when the page is linked to LRU but page_cgroup isn't */
+       if (PageLRU(page) && list_empty(&pc->lru))
+               mem_cgroup_add_lru_list(page, page_lru(page));
+       spin_unlock_irqrestore(&zone->lru_lock, flags);
 }
 
+
 void mem_cgroup_move_lists(struct page *page,
                           enum lru_list from, enum lru_list to)
 {
@@ -379,31 +490,24 @@ void mem_cgroup_move_lists(struct page *page,
 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
 {
        int ret;
+       struct mem_cgroup *curr = NULL;
 
        task_lock(task);
-       ret = task->mm && mm_match_cgroup(task->mm, mem);
+       rcu_read_lock();
+       curr = try_get_mem_cgroup_from_mm(task->mm);
+       rcu_read_unlock();
        task_unlock(task);
+       if (!curr)
+               return 0;
+       if (curr->use_hierarchy)
+               ret = css_is_ancestor(&curr->css, &mem->css);
+       else
+               ret = (curr == mem);
+       css_put(&curr->css);
        return ret;
 }
 
 /*
- * Calculate mapped_ratio under memory controller. This will be used in
- * vmscan.c for deteremining we have to reclaim mapped pages.
- */
-int mem_cgroup_calc_mapped_ratio(struct mem_cgroup *mem)
-{
-       long total, rss;
-
-       /*
-        * usage is recorded in bytes. But, here, we assume the number of
-        * physical pages can be represented by "long" on any arch.
-        */
-       total = (long) (mem->res.usage >> PAGE_SHIFT) + 1L;
-       rss = (long)mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
-       return (int)((rss * 100L) / total);
-}
-
-/*
  * prev_priority control...this will be used in memory reclaim path.
  */
 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
@@ -432,15 +536,43 @@ void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
        spin_unlock(&mem->reclaim_param_lock);
 }
 
-int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
+static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
+{
+       unsigned long active;
+       unsigned long inactive;
+       unsigned long gb;
+       unsigned long inactive_ratio;
+
+       inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
+       active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
+
+       gb = (inactive + active) >> (30 - PAGE_SHIFT);
+       if (gb)
+               inactive_ratio = int_sqrt(10 * gb);
+       else
+               inactive_ratio = 1;
+
+       if (present_pages) {
+               present_pages[0] = inactive;
+               present_pages[1] = active;
+       }
+
+       return inactive_ratio;
+}
+
+int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
 {
        unsigned long active;
        unsigned long inactive;
+       unsigned long present_pages[2];
+       unsigned long inactive_ratio;
+
+       inactive_ratio = calc_inactive_ratio(memcg, present_pages);
 
-       inactive = mem_cgroup_get_all_zonestat(memcg, LRU_INACTIVE_ANON);
-       active = mem_cgroup_get_all_zonestat(memcg, LRU_ACTIVE_ANON);
+       inactive = present_pages[0];
+       active = present_pages[1];
 
-       if (inactive * memcg->inactive_ratio < active)
+       if (inactive * inactive_ratio < active)
                return 1;
 
        return 0;
@@ -477,6 +609,14 @@ mem_cgroup_get_reclaim_stat_from_page(struct page *page)
                return NULL;
 
        pc = lookup_page_cgroup(page);
+       /*
+        * Used bit is set without atomic ops but after smp_wmb().
+        * For making pc->mem_cgroup visible, insert smp_rmb() here.
+        */
+       smp_rmb();
+       if (!PageCgroupUsed(pc))
+               return NULL;
+
        mz = page_cgroup_zoneinfo(pc);
        if (!mz)
                return NULL;
@@ -531,59 +671,117 @@ unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
 #define mem_cgroup_from_res_counter(counter, member)   \
        container_of(counter, struct mem_cgroup, member)
 
-/*
- * This routine finds the DFS walk successor. This routine should be
- * called with cgroup_mutex held
+static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
+{
+       if (do_swap_account) {
+               if (res_counter_check_under_limit(&mem->res) &&
+                       res_counter_check_under_limit(&mem->memsw))
+                       return true;
+       } else
+               if (res_counter_check_under_limit(&mem->res))
+                       return true;
+       return false;
+}
+
+static unsigned int get_swappiness(struct mem_cgroup *memcg)
+{
+       struct cgroup *cgrp = memcg->css.cgroup;
+       unsigned int swappiness;
+
+       /* root ? */
+       if (cgrp->parent == NULL)
+               return vm_swappiness;
+
+       spin_lock(&memcg->reclaim_param_lock);
+       swappiness = memcg->swappiness;
+       spin_unlock(&memcg->reclaim_param_lock);
+
+       return swappiness;
+}
+
+static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data)
+{
+       int *val = data;
+       (*val)++;
+       return 0;
+}
+
+/**
+ * mem_cgroup_print_mem_info: Called from OOM with tasklist_lock held in read mode.
+ * @memcg: The memory cgroup that went over limit
+ * @p: Task that is going to be killed
+ *
+ * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
+ * enabled
  */
-static struct mem_cgroup *
-mem_cgroup_get_next_node(struct mem_cgroup *curr, struct mem_cgroup *root_mem)
+void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
 {
-       struct cgroup *cgroup, *curr_cgroup, *root_cgroup;
+       struct cgroup *task_cgrp;
+       struct cgroup *mem_cgrp;
+       /*
+        * Need a buffer in BSS, can't rely on allocations. The code relies
+        * on the assumption that OOM is serialized for memory controller.
+        * If this assumption is broken, revisit this code.
+        */
+       static char memcg_name[PATH_MAX];
+       int ret;
 
-       curr_cgroup = curr->css.cgroup;
-       root_cgroup = root_mem->css.cgroup;
+       if (!memcg)
+               return;
+
+
+       rcu_read_lock();
 
-       if (!list_empty(&curr_cgroup->children)) {
+       mem_cgrp = memcg->css.cgroup;
+       task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
+
+       ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
+       if (ret < 0) {
                /*
-                * Walk down to children
+                * Unfortunately, we are unable to convert to a useful name
+                * But we'll still print out the usage information
                 */
-               mem_cgroup_put(curr);
-               cgroup = list_entry(curr_cgroup->children.next,
-                                               struct cgroup, sibling);
-               curr = mem_cgroup_from_cont(cgroup);
-               mem_cgroup_get(curr);
+               rcu_read_unlock();
                goto done;
        }
+       rcu_read_unlock();
 
-visit_parent:
-       if (curr_cgroup == root_cgroup) {
-               mem_cgroup_put(curr);
-               curr = root_mem;
-               mem_cgroup_get(curr);
-               goto done;
-       }
+       printk(KERN_INFO "Task in %s killed", memcg_name);
 
-       /*
-        * Goto next sibling
-        */
-       if (curr_cgroup->sibling.next != &curr_cgroup->parent->children) {
-               mem_cgroup_put(curr);
-               cgroup = list_entry(curr_cgroup->sibling.next, struct cgroup,
-                                               sibling);
-               curr = mem_cgroup_from_cont(cgroup);
-               mem_cgroup_get(curr);
+       rcu_read_lock();
+       ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
+       if (ret < 0) {
+               rcu_read_unlock();
                goto done;
        }
+       rcu_read_unlock();
 
        /*
-        * Go up to next parent and next parent's sibling if need be
+        * Continues from above, so we don't need an KERN_ level
         */
-       curr_cgroup = curr_cgroup->parent;
-       goto visit_parent;
-
+       printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
 done:
-       root_mem->last_scanned_child = curr;
-       return curr;
+
+       printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
+               res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
+               res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
+               res_counter_read_u64(&memcg->res, RES_FAILCNT));
+       printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
+               "failcnt %llu\n",
+               res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
+               res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
+               res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
+}
+
+/*
+ * This function returns the number of memcg under hierarchy tree. Returns
+ * 1(self count) if no children.
+ */
+static int mem_cgroup_count_children(struct mem_cgroup *mem)
+{
+       int num = 0;
+       mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb);
+       return num;
 }
 
 /*
@@ -592,115 +790,83 @@ done:
  * that to reclaim free pages from.
  */
 static struct mem_cgroup *
-mem_cgroup_get_first_node(struct mem_cgroup *root_mem)
+mem_cgroup_select_victim(struct mem_cgroup *root_mem)
 {
-       struct cgroup *cgroup;
-       struct mem_cgroup *ret;
-       bool obsolete = (root_mem->last_scanned_child &&
-                               root_mem->last_scanned_child->obsolete);
+       struct mem_cgroup *ret = NULL;
+       struct cgroup_subsys_state *css;
+       int nextid, found;
 
-       /*
-        * Scan all children under the mem_cgroup mem
-        */
-       cgroup_lock();
-       if (list_empty(&root_mem->css.cgroup->children)) {
+       if (!root_mem->use_hierarchy) {
+               css_get(&root_mem->css);
                ret = root_mem;
-               goto done;
        }
 
-       if (!root_mem->last_scanned_child || obsolete) {
-
-               if (obsolete)
-                       mem_cgroup_put(root_mem->last_scanned_child);
+       while (!ret) {
+               rcu_read_lock();
+               nextid = root_mem->last_scanned_child + 1;
+               css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
+                                  &found);
+               if (css && css_tryget(css))
+                       ret = container_of(css, struct mem_cgroup, css);
 
-               cgroup = list_first_entry(&root_mem->css.cgroup->children,
-                               struct cgroup, sibling);
-               ret = mem_cgroup_from_cont(cgroup);
-               mem_cgroup_get(ret);
-       } else
-               ret = mem_cgroup_get_next_node(root_mem->last_scanned_child,
-                                               root_mem);
+               rcu_read_unlock();
+               /* Updates scanning parameter */
+               spin_lock(&root_mem->reclaim_param_lock);
+               if (!css) {
+                       /* this means start scan from ID:1 */
+                       root_mem->last_scanned_child = 0;
+               } else
+                       root_mem->last_scanned_child = found;
+               spin_unlock(&root_mem->reclaim_param_lock);
+       }
 
-done:
-       root_mem->last_scanned_child = ret;
-       cgroup_unlock();
        return ret;
 }
 
-static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
-{
-       if (do_swap_account) {
-               if (res_counter_check_under_limit(&mem->res) &&
-                       res_counter_check_under_limit(&mem->memsw))
-                       return true;
-       } else
-               if (res_counter_check_under_limit(&mem->res))
-                       return true;
-       return false;
-}
-
-static unsigned int get_swappiness(struct mem_cgroup *memcg)
-{
-       struct cgroup *cgrp = memcg->css.cgroup;
-       unsigned int swappiness;
-
-       /* root ? */
-       if (cgrp->parent == NULL)
-               return vm_swappiness;
-
-       spin_lock(&memcg->reclaim_param_lock);
-       swappiness = memcg->swappiness;
-       spin_unlock(&memcg->reclaim_param_lock);
-
-       return swappiness;
-}
-
 /*
- * Dance down the hierarchy if needed to reclaim memory. We remember the
- * last child we reclaimed from, so that we don't end up penalizing
- * one child extensively based on its position in the children list.
+ * Scan the hierarchy if needed to reclaim memory. We remember the last child
+ * we reclaimed from, so that we don't end up penalizing one child extensively
+ * based on its position in the children list.
  *
  * root_mem is the original ancestor that we've been reclaim from.
+ *
+ * We give up and return to the caller when we visit root_mem twice.
+ * (other groups can be removed while we're walking....)
+ *
+ * If shrink==true, for avoiding to free too much, this returns immedieately.
  */
 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
-                                               gfp_t gfp_mask, bool noswap)
-{
-       struct mem_cgroup *next_mem;
-       int ret = 0;
-
-       /*
-        * Reclaim unconditionally and don't check for return value.
-        * We need to reclaim in the current group and down the tree.
-        * One might think about checking for children before reclaiming,
-        * but there might be left over accounting, even after children
-        * have left.
-        */
-       ret = try_to_free_mem_cgroup_pages(root_mem, gfp_mask, noswap,
-                                          get_swappiness(root_mem));
-       if (mem_cgroup_check_under_limit(root_mem))
-               return 0;
-       if (!root_mem->use_hierarchy)
-               return ret;
-
-       next_mem = mem_cgroup_get_first_node(root_mem);
-
-       while (next_mem != root_mem) {
-               if (next_mem->obsolete) {
-                       mem_cgroup_put(next_mem);
-                       cgroup_lock();
-                       next_mem = mem_cgroup_get_first_node(root_mem);
-                       cgroup_unlock();
+                                  gfp_t gfp_mask, bool noswap, bool shrink)
+{
+       struct mem_cgroup *victim;
+       int ret, total = 0;
+       int loop = 0;
+
+       while (loop < 2) {
+               victim = mem_cgroup_select_victim(root_mem);
+               if (victim == root_mem)
+                       loop++;
+               if (!mem_cgroup_local_usage(&victim->stat)) {
+                       /* this cgroup's local usage == 0 */
+                       css_put(&victim->css);
                        continue;
                }
-               ret = try_to_free_mem_cgroup_pages(next_mem, gfp_mask, noswap,
-                                                  get_swappiness(next_mem));
+               /* we use swappiness of local cgroup */
+               ret = try_to_free_mem_cgroup_pages(victim, gfp_mask, noswap,
+                                                  get_swappiness(victim));
+               css_put(&victim->css);
+               /*
+                * At shrinking usage, we can't check we should stop here or
+                * reclaim more. It's depends on callers. last_scanned_child
+                * will work enough for keeping fairness under tree.
+                */
+               if (shrink)
+                       return ret;
+               total += ret;
                if (mem_cgroup_check_under_limit(root_mem))
-                       return 0;
-               cgroup_lock();
-               next_mem = mem_cgroup_get_next_node(next_mem, root_mem);
-               cgroup_unlock();
+                       return 1 + total;
        }
-       return ret;
+       return total;
 }
 
 bool mem_cgroup_oom_called(struct task_struct *task)
@@ -719,6 +885,19 @@ bool mem_cgroup_oom_called(struct task_struct *task)
        rcu_read_unlock();
        return ret;
 }
+
+static int record_last_oom_cb(struct mem_cgroup *mem, void *data)
+{
+       mem->last_oom_jiffies = jiffies;
+       return 0;
+}
+
+static void record_last_oom(struct mem_cgroup *mem)
+{
+       mem_cgroup_walk_tree(mem, NULL, record_last_oom_cb);
+}
+
+
 /*
  * Unlike exported interface, "oom" parameter is added. if oom==true,
  * oom-killer can be invoked.
@@ -743,23 +922,17 @@ static int __mem_cgroup_try_charge(struct mm_struct *mm,
         * thread group leader migrates. It's possible that mm is not
         * set, if so charge the init_mm (happens for pagecache usage).
         */
-       if (likely(!*memcg)) {
-               rcu_read_lock();
-               mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
-               if (unlikely(!mem)) {
-                       rcu_read_unlock();
-                       return 0;
-               }
-               /*
-                * For every charge from the cgroup, increment reference count
-                */
-               css_get(&mem->css);
+       mem = *memcg;
+       if (likely(!mem)) {
+               mem = try_get_mem_cgroup_from_mm(mm);
                *memcg = mem;
-               rcu_read_unlock();
        } else {
-               mem = *memcg;
                css_get(&mem->css);
        }
+       if (unlikely(!mem))
+               return 0;
+
+       VM_BUG_ON(mem_cgroup_is_obsolete(mem));
 
        while (1) {
                int ret;
@@ -787,7 +960,9 @@ static int __mem_cgroup_try_charge(struct mm_struct *mm,
                        goto nomem;
 
                ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, gfp_mask,
-                                                       noswap);
+                                                       noswap, false);
+               if (ret)
+                       continue;
 
                /*
                 * try_to_free_mem_cgroup_pages() might not give us a full
@@ -802,8 +977,10 @@ static int __mem_cgroup_try_charge(struct mm_struct *mm,
 
                if (!nr_retries--) {
                        if (oom) {
+                               mutex_lock(&memcg_tasklist);
                                mem_cgroup_out_of_memory(mem_over_limit, gfp_mask);
-                               mem_over_limit->last_oom_jiffies = jiffies;
+                               mutex_unlock(&memcg_tasklist);
+                               record_last_oom(mem_over_limit);
                        }
                        goto nomem;
                }
@@ -814,27 +991,36 @@ nomem:
        return -ENOMEM;
 }
 
-/**
- * mem_cgroup_try_charge - get charge of PAGE_SIZE.
- * @mm: an mm_struct which is charged against. (when *memcg is NULL)
- * @gfp_mask: gfp_mask for reclaim.
- * @memcg: a pointer to memory cgroup which is charged against.
- *
- * charge against memory cgroup pointed by *memcg. if *memcg == NULL, estimated
- * memory cgroup from @mm is got and stored in *memcg.
- *
- * Returns 0 if success. -ENOMEM at failure.
- * This call can invoke OOM-Killer.
- */
-
-int mem_cgroup_try_charge(struct mm_struct *mm,
-                         gfp_t mask, struct mem_cgroup **memcg)
+static struct mem_cgroup *try_get_mem_cgroup_from_swapcache(struct page *page)
 {
-       return __mem_cgroup_try_charge(mm, mask, memcg, true);
+       struct mem_cgroup *mem;
+       struct page_cgroup *pc;
+       swp_entry_t ent;
+
+       VM_BUG_ON(!PageLocked(page));
+
+       if (!PageSwapCache(page))
+               return NULL;
+
+       pc = lookup_page_cgroup(page);
+       /*
+        * Used bit of swapcache is solid under page lock.
+        */
+       if (PageCgroupUsed(pc))
+               mem = pc->mem_cgroup;
+       else {
+               ent.val = page_private(page);
+               mem = lookup_swap_cgroup(ent);
+       }
+       if (!mem)
+               return NULL;
+       if (!css_tryget(&mem->css))
+               return NULL;
+       return mem;
 }
 
 /*
- * commit a charge got by mem_cgroup_try_charge() and makes page_cgroup to be
+ * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
  * USED state. If already USED, uncharge and return.
  */
 
@@ -904,14 +1090,15 @@ static int mem_cgroup_move_account(struct page_cgroup *pc,
        if (pc->mem_cgroup != from)
                goto out;
 
-       css_put(&from->css);
        res_counter_uncharge(&from->res, PAGE_SIZE);
        mem_cgroup_charge_statistics(from, pc, false);
        if (do_swap_account)
                res_counter_uncharge(&from->memsw, PAGE_SIZE);
+       css_put(&from->css);
+
+       css_get(&to->css);
        pc->mem_cgroup = to;
        mem_cgroup_charge_statistics(to, pc, true);
-       css_get(&to->css);
        ret = 0;
 out:
        unlock_page_cgroup(pc);
@@ -944,8 +1131,10 @@ static int mem_cgroup_move_parent(struct page_cgroup *pc,
        if (ret || !parent)
                return ret;
 
-       if (!get_page_unless_zero(page))
-               return -EBUSY;
+       if (!get_page_unless_zero(page)) {
+               ret = -EBUSY;
+               goto uncharge;
+       }
 
        ret = isolate_lru_page(page);
 
@@ -954,19 +1143,23 @@ static int mem_cgroup_move_parent(struct page_cgroup *pc,
 
        ret = mem_cgroup_move_account(pc, child, parent);
 
-       /* drop extra refcnt by try_charge() (move_account increment one) */
-       css_put(&parent->css);
        putback_lru_page(page);
        if (!ret) {
                put_page(page);
+               /* drop extra refcnt by try_charge() */
+               css_put(&parent->css);
                return 0;
        }
-       /* uncharge if move fails */
+
 cancel:
+       put_page(page);
+uncharge:
+       /* drop extra refcnt by try_charge() */
+       css_put(&parent->css);
+       /* uncharge if move fails */
        res_counter_uncharge(&parent->res, PAGE_SIZE);
        if (do_swap_account)
                res_counter_uncharge(&parent->memsw, PAGE_SIZE);
-       put_page(page);
        return ret;
 }
 
@@ -1024,6 +1217,9 @@ int mem_cgroup_newpage_charge(struct page *page,
 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
                                gfp_t gfp_mask)
 {
+       struct mem_cgroup *mem = NULL;
+       int ret;
+
        if (mem_cgroup_disabled())
                return 0;
        if (PageCompound(page))
@@ -1036,6 +1232,8 @@ int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
         * For GFP_NOWAIT case, the page may be pre-charged before calling
         * add_to_page_cache(). (See shmem.c) check it here and avoid to call
         * charge twice. (It works but has to pay a bit larger cost.)
+        * And when the page is SwapCache, it should take swap information
+        * into account. This is under lock_page() now.
         */
        if (!(gfp_mask & __GFP_WAIT)) {
                struct page_cgroup *pc;
@@ -1052,30 +1250,60 @@ int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
                unlock_page_cgroup(pc);
        }
 
-       if (unlikely(!mm))
+       if (do_swap_account && PageSwapCache(page)) {
+               mem = try_get_mem_cgroup_from_swapcache(page);
+               if (mem)
+                       mm = NULL;
+                 else
+                       mem = NULL;
+               /* SwapCache may be still linked to LRU now. */
+               mem_cgroup_lru_del_before_commit_swapcache(page);
+       }
+
+       if (unlikely(!mm && !mem))
                mm = &init_mm;
 
        if (page_is_file_cache(page))
                return mem_cgroup_charge_common(page, mm, gfp_mask,
                                MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
-       else
-               return mem_cgroup_charge_common(page, mm, gfp_mask,
-                               MEM_CGROUP_CHARGE_TYPE_SHMEM, NULL);
+
+       ret = mem_cgroup_charge_common(page, mm, gfp_mask,
+                               MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
+       if (mem)
+               css_put(&mem->css);
+       if (PageSwapCache(page))
+               mem_cgroup_lru_add_after_commit_swapcache(page);
+
+       if (do_swap_account && !ret && PageSwapCache(page)) {
+               swp_entry_t ent = {.val = page_private(page)};
+               /* avoid double counting */
+               mem = swap_cgroup_record(ent, NULL);
+               if (mem) {
+                       res_counter_uncharge(&mem->memsw, PAGE_SIZE);
+                       mem_cgroup_put(mem);
+               }
+       }
+       return ret;
 }
 
+/*
+ * While swap-in, try_charge -> commit or cancel, the page is locked.
+ * And when try_charge() successfully returns, one refcnt to memcg without
+ * struct page_cgroup is aquired. This refcnt will be cumsumed by
+ * "commit()" or removed by "cancel()"
+ */
 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
                                 struct page *page,
                                 gfp_t mask, struct mem_cgroup **ptr)
 {
        struct mem_cgroup *mem;
-       swp_entry_t     ent;
+       int ret;
 
        if (mem_cgroup_disabled())
                return 0;
 
        if (!do_swap_account)
                goto charge_cur_mm;
-
        /*
         * A racing thread's fault, or swapoff, may have already updated
         * the pte, and even removed page from swap cache: return success
@@ -1083,70 +1311,20 @@ int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
         */
        if (!PageSwapCache(page))
                return 0;
-
-       ent.val = page_private(page);
-
-       mem = lookup_swap_cgroup(ent);
-       if (!mem || mem->obsolete)
+       mem = try_get_mem_cgroup_from_swapcache(page);
+       if (!mem)
                goto charge_cur_mm;
        *ptr = mem;
-       return __mem_cgroup_try_charge(NULL, mask, ptr, true);
+       ret = __mem_cgroup_try_charge(NULL, mask, ptr, true);
+       /* drop extra refcnt from tryget */
+       css_put(&mem->css);
+       return ret;
 charge_cur_mm:
        if (unlikely(!mm))
                mm = &init_mm;
        return __mem_cgroup_try_charge(mm, mask, ptr, true);
 }
 
-#ifdef CONFIG_SWAP
-
-int mem_cgroup_cache_charge_swapin(struct page *page,
-                       struct mm_struct *mm, gfp_t mask, bool locked)
-{
-       int ret = 0;
-
-       if (mem_cgroup_disabled())
-               return 0;
-       if (unlikely(!mm))
-               mm = &init_mm;
-       if (!locked)
-               lock_page(page);
-       /*
-        * If not locked, the page can be dropped from SwapCache until
-        * we reach here.
-        */
-       if (PageSwapCache(page)) {
-               struct mem_cgroup *mem = NULL;
-               swp_entry_t ent;
-
-               ent.val = page_private(page);
-               if (do_swap_account) {
-                       mem = lookup_swap_cgroup(ent);
-                       if (mem && mem->obsolete)
-                               mem = NULL;
-                       if (mem)
-                               mm = NULL;
-               }
-               ret = mem_cgroup_charge_common(page, mm, mask,
-                               MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
-
-               if (!ret && do_swap_account) {
-                       /* avoid double counting */
-                       mem = swap_cgroup_record(ent, NULL);
-                       if (mem) {
-                               res_counter_uncharge(&mem->memsw, PAGE_SIZE);
-                               mem_cgroup_put(mem);
-                       }
-               }
-       }
-       if (!locked)
-               unlock_page(page);
-       /* add this page(page_cgroup) to the LRU we want. */
-       mem_cgroup_lru_fixup(page);
-
-       return ret;
-}
-#endif
-
 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
 {
        struct page_cgroup *pc;
@@ -1156,26 +1334,28 @@ void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
        if (!ptr)
                return;
        pc = lookup_page_cgroup(page);
+       mem_cgroup_lru_del_before_commit_swapcache(page);
        __mem_cgroup_commit_charge(ptr, pc, MEM_CGROUP_CHARGE_TYPE_MAPPED);
+       mem_cgroup_lru_add_after_commit_swapcache(page);
        /*
         * Now swap is on-memory. This means this page may be
         * counted both as mem and swap....double count.
-        * Fix it by uncharging from memsw. This SwapCache is stable
-        * because we're still under lock_page().
+        * Fix it by uncharging from memsw. Basically, this SwapCache is stable
+        * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
+        * may call delete_from_swap_cache() before reach here.
         */
-       if (do_swap_account) {
+       if (do_swap_account && PageSwapCache(page)) {
                swp_entry_t ent = {.val = page_private(page)};
                struct mem_cgroup *memcg;
                memcg = swap_cgroup_record(ent, NULL);
                if (memcg) {
-                       /* If memcg is obsolete, memcg can be != ptr */
                        res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
                        mem_cgroup_put(memcg);
                }
 
        }
        /* add this page(page_cgroup) to the LRU we want. */
-       mem_cgroup_lru_fixup(page);
+
 }
 
 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
@@ -1240,9 +1420,15 @@ __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
        res_counter_uncharge(&mem->res, PAGE_SIZE);
        if (do_swap_account && (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT))
                res_counter_uncharge(&mem->memsw, PAGE_SIZE);
-
        mem_cgroup_charge_statistics(mem, pc, false);
+
        ClearPageCgroupUsed(pc);
+       /*
+        * pc->mem_cgroup is not cleared here. It will be accessed when it's
+        * freed from LRU. This is safe because uncharged page is expected not
+        * to be reused (freed soon). Exception is SwapCache, it's handled by
+        * special functions.
+        */
 
        mz = page_cgroup_zoneinfo(pc);
        unlock_page_cgroup(pc);
@@ -1336,7 +1522,7 @@ int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
        unlock_page_cgroup(pc);
 
        if (mem) {
-               ret = mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem);
+               ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
                css_put(&mem->css);
        }
        *ptr = mem;
@@ -1400,29 +1586,26 @@ void mem_cgroup_end_migration(struct mem_cgroup *mem,
  * This is typically used for page reclaiming for shmem for reducing side
  * effect of page allocation from shmem, which is used by some mem_cgroup.
  */
-int mem_cgroup_shrink_usage(struct mm_struct *mm, gfp_t gfp_mask)
+int mem_cgroup_shrink_usage(struct page *page,
+                           struct mm_struct *mm,
+                           gfp_t gfp_mask)
 {
-       struct mem_cgroup *mem;
+       struct mem_cgroup *mem = NULL;
        int progress = 0;
        int retry = MEM_CGROUP_RECLAIM_RETRIES;
 
        if (mem_cgroup_disabled())
                return 0;
-       if (!mm)
+       if (page)
+               mem = try_get_mem_cgroup_from_swapcache(page);
+       if (!mem && mm)
+               mem = try_get_mem_cgroup_from_mm(mm);
+       if (unlikely(!mem))
                return 0;
 
-       rcu_read_lock();
-       mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
-       if (unlikely(!mem)) {
-               rcu_read_unlock();
-               return 0;
-       }
-       css_get(&mem->css);
-       rcu_read_unlock();
-
        do {
-               progress = try_to_free_mem_cgroup_pages(mem, gfp_mask, true,
-                                                       get_swappiness(mem));
+               progress = mem_cgroup_hierarchical_reclaim(mem,
+                                       gfp_mask, true, false);
                progress += mem_cgroup_check_under_limit(mem);
        } while (!progress && --retry);
 
@@ -1432,39 +1615,26 @@ int mem_cgroup_shrink_usage(struct mm_struct *mm, gfp_t gfp_mask)
        return 0;
 }
 
-/*
- * The inactive anon list should be small enough that the VM never has to
- * do too much work, but large enough that each inactive page has a chance
- * to be referenced again before it is swapped out.
- *
- * this calculation is straightforward porting from
- * page_alloc.c::setup_per_zone_inactive_ratio().
- * it describe more detail.
- */
-static void mem_cgroup_set_inactive_ratio(struct mem_cgroup *memcg)
-{
-       unsigned int gb, ratio;
-
-       gb = res_counter_read_u64(&memcg->res, RES_LIMIT) >> 30;
-       if (gb)
-               ratio = int_sqrt(10 * gb);
-       else
-               ratio = 1;
-
-       memcg->inactive_ratio = ratio;
-
-}
-
 static DEFINE_MUTEX(set_limit_mutex);
 
 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
                                unsigned long long val)
 {
-
-       int retry_count = MEM_CGROUP_RECLAIM_RETRIES;
+       int retry_count;
        int progress;
        u64 memswlimit;
        int ret = 0;
+       int children = mem_cgroup_count_children(memcg);
+       u64 curusage, oldusage;
+
+       /*
+        * For keeping hierarchical_reclaim simple, how long we should retry
+        * is depends on callers. We set our retry-count to be function
+        * of # of children which we should visit in this loop.
+        */
+       retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
+
+       oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
 
        while (retry_count) {
                if (signal_pending(current)) {
@@ -1489,29 +1659,32 @@ static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
                if (!ret)
                        break;
 
-               progress = try_to_free_mem_cgroup_pages(memcg,
-                                                       GFP_KERNEL,
-                                                       false,
-                                                       get_swappiness(memcg));
-               if (!progress)                  retry_count--;
+               progress = mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL,
+                                                  false, true);
+               curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
+               /* Usage is reduced ? */
+               if (curusage >= oldusage)
+                       retry_count--;
+               else
+                       oldusage = curusage;
        }
 
-       if (!ret)
-               mem_cgroup_set_inactive_ratio(memcg);
-
        return ret;
 }
 
 int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
                                unsigned long long val)
 {
-       int retry_count = MEM_CGROUP_RECLAIM_RETRIES;
+       int retry_count;
        u64 memlimit, oldusage, curusage;
-       int ret;
+       int children = mem_cgroup_count_children(memcg);
+       int ret = -EBUSY;
 
        if (!do_swap_account)
                return -EINVAL;
-
+       /* see mem_cgroup_resize_res_limit */
+       retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
+       oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
        while (retry_count) {
                if (signal_pending(current)) {
                        ret = -EINTR;
@@ -1535,12 +1708,13 @@ int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
                if (!ret)
                        break;
 
-               oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
-               try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL, true,
-                                            get_swappiness(memcg));
+               mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL, true, true);
                curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
+               /* Usage is reduced ? */
                if (curusage >= oldusage)
                        retry_count--;
+               else
+                       oldusage = curusage;
        }
        return ret;
 }
@@ -1628,7 +1802,7 @@ move_account:
                /* This is for making all *used* pages to be on LRU. */
                lru_add_drain_all();
                ret = 0;
-               for_each_node_state(node, N_POSSIBLE) {
+               for_each_node_state(node, N_HIGH_MEMORY) {
                        for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
                                enum lru_list l;
                                for_each_lru(l) {
@@ -1783,6 +1957,34 @@ static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
        return ret;
 }
 
+static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
+               unsigned long long *mem_limit, unsigned long long *memsw_limit)
+{
+       struct cgroup *cgroup;
+       unsigned long long min_limit, min_memsw_limit, tmp;
+
+       min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
+       min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
+       cgroup = memcg->css.cgroup;
+       if (!memcg->use_hierarchy)
+               goto out;
+
+       while (cgroup->parent) {
+               cgroup = cgroup->parent;
+               memcg = mem_cgroup_from_cont(cgroup);
+               if (!memcg->use_hierarchy)
+                       break;
+               tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
+               min_limit = min(min_limit, tmp);
+               tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
+               min_memsw_limit = min(min_memsw_limit, tmp);
+       }
+out:
+       *mem_limit = min_limit;
+       *memsw_limit = min_memsw_limit;
+       return;
+}
+
 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
 {
        struct mem_cgroup *mem;
@@ -1808,57 +2010,106 @@ static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
        return 0;
 }
 
-static const struct mem_cgroup_stat_desc {
-       const char *msg;
-       u64 unit;
-} mem_cgroup_stat_desc[] = {
-       [MEM_CGROUP_STAT_CACHE] = { "cache", PAGE_SIZE, },
-       [MEM_CGROUP_STAT_RSS] = { "rss", PAGE_SIZE, },
-       [MEM_CGROUP_STAT_PGPGIN_COUNT] = {"pgpgin", 1, },
-       [MEM_CGROUP_STAT_PGPGOUT_COUNT] = {"pgpgout", 1, },
+
+/* For read statistics */
+enum {
+       MCS_CACHE,
+       MCS_RSS,
+       MCS_PGPGIN,
+       MCS_PGPGOUT,
+       MCS_INACTIVE_ANON,
+       MCS_ACTIVE_ANON,
+       MCS_INACTIVE_FILE,
+       MCS_ACTIVE_FILE,
+       MCS_UNEVICTABLE,
+       NR_MCS_STAT,
+};
+
+struct mcs_total_stat {
+       s64 stat[NR_MCS_STAT];
+};
+
+struct {
+       char *local_name;
+       char *total_name;
+} memcg_stat_strings[NR_MCS_STAT] = {
+       {"cache", "total_cache"},
+       {"rss", "total_rss"},
+       {"pgpgin", "total_pgpgin"},
+       {"pgpgout", "total_pgpgout"},
+       {"inactive_anon", "total_inactive_anon"},
+       {"active_anon", "total_active_anon"},
+       {"inactive_file", "total_inactive_file"},
+       {"active_file", "total_active_file"},
+       {"unevictable", "total_unevictable"}
 };
 
+
+static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data)
+{
+       struct mcs_total_stat *s = data;
+       s64 val;
+
+       /* per cpu stat */
+       val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_CACHE);
+       s->stat[MCS_CACHE] += val * PAGE_SIZE;
+       val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
+       s->stat[MCS_RSS] += val * PAGE_SIZE;
+       val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGIN_COUNT);
+       s->stat[MCS_PGPGIN] += val;
+       val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGOUT_COUNT);
+       s->stat[MCS_PGPGOUT] += val;
+
+       /* per zone stat */
+       val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
+       s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
+       val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
+       s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
+       val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
+       s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
+       val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
+       s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
+       val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
+       s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
+       return 0;
+}
+
+static void
+mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
+{
+       mem_cgroup_walk_tree(mem, s, mem_cgroup_get_local_stat);
+}
+
 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
                                 struct cgroup_map_cb *cb)
 {
        struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
-       struct mem_cgroup_stat *stat = &mem_cont->stat;
+       struct mcs_total_stat mystat;
        int i;
 
-       for (i = 0; i < ARRAY_SIZE(stat->cpustat[0].count); i++) {
-               s64 val;
+       memset(&mystat, 0, sizeof(mystat));
+       mem_cgroup_get_local_stat(mem_cont, &mystat);
 
-               val = mem_cgroup_read_stat(stat, i);
-               val *= mem_cgroup_stat_desc[i].unit;
-               cb->fill(cb, mem_cgroup_stat_desc[i].msg, val);
-       }
-       /* showing # of active pages */
-       {
-               unsigned long active_anon, inactive_anon;
-               unsigned long active_file, inactive_file;
-               unsigned long unevictable;
-
-               inactive_anon = mem_cgroup_get_all_zonestat(mem_cont,
-                                               LRU_INACTIVE_ANON);
-               active_anon = mem_cgroup_get_all_zonestat(mem_cont,
-                                               LRU_ACTIVE_ANON);
-               inactive_file = mem_cgroup_get_all_zonestat(mem_cont,
-                                               LRU_INACTIVE_FILE);
-               active_file = mem_cgroup_get_all_zonestat(mem_cont,
-                                               LRU_ACTIVE_FILE);
-               unevictable = mem_cgroup_get_all_zonestat(mem_cont,
-                                                       LRU_UNEVICTABLE);
-
-               cb->fill(cb, "active_anon", (active_anon) * PAGE_SIZE);
-               cb->fill(cb, "inactive_anon", (inactive_anon) * PAGE_SIZE);
-               cb->fill(cb, "active_file", (active_file) * PAGE_SIZE);
-               cb->fill(cb, "inactive_file", (inactive_file) * PAGE_SIZE);
-               cb->fill(cb, "unevictable", unevictable * PAGE_SIZE);
+       for (i = 0; i < NR_MCS_STAT; i++)
+               cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
 
+       /* Hierarchical information */
+       {
+               unsigned long long limit, memsw_limit;
+               memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
+               cb->fill(cb, "hierarchical_memory_limit", limit);
+               if (do_swap_account)
+                       cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
        }
 
+       memset(&mystat, 0, sizeof(mystat));
+       mem_cgroup_get_total_stat(mem_cont, &mystat);
+       for (i = 0; i < NR_MCS_STAT; i++)
+               cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
+
+
 #ifdef CONFIG_DEBUG_VM
-       cb->fill(cb, "inactive_ratio", mem_cont->inactive_ratio);
+       cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
 
        {
                int nid, zid;
@@ -1901,6 +2152,7 @@ static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
 {
        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
        struct mem_cgroup *parent;
+
        if (val > 100)
                return -EINVAL;
 
@@ -1908,15 +2160,22 @@ static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
                return -EINVAL;
 
        parent = mem_cgroup_from_cont(cgrp->parent);
+
+       cgroup_lock();
+
        /* If under hierarchy, only empty-root can set this value */
        if ((parent->use_hierarchy) ||
-           (memcg->use_hierarchy && !list_empty(&cgrp->children)))
+           (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
+               cgroup_unlock();
                return -EINVAL;
+       }
 
        spin_lock(&memcg->reclaim_param_lock);
        memcg->swappiness = val;
        spin_unlock(&memcg->reclaim_param_lock);
 
+       cgroup_unlock();
+
        return 0;
 }
 
@@ -2071,19 +2330,14 @@ static struct mem_cgroup *mem_cgroup_alloc(void)
  * the number of reference from swap_cgroup and free mem_cgroup when
  * it goes down to 0.
  *
- * When mem_cgroup is destroyed, mem->obsolete will be set to 0 and
- * entry which points to this memcg will be ignore at swapin.
- *
  * Removal of cgroup itself succeeds regardless of refs from swap.
  */
 
-static void mem_cgroup_free(struct mem_cgroup *mem)
+static void __mem_cgroup_free(struct mem_cgroup *mem)
 {
        int node;
 
-       if (atomic_read(&mem->refcnt) > 0)
-               return;
-
+       free_css_id(&mem_cgroup_subsys, &mem->css);
 
        for_each_node_state(node, N_POSSIBLE)
                free_mem_cgroup_per_zone_info(mem, node);
@@ -2102,12 +2356,22 @@ static void mem_cgroup_get(struct mem_cgroup *mem)
 static void mem_cgroup_put(struct mem_cgroup *mem)
 {
        if (atomic_dec_and_test(&mem->refcnt)) {
-               if (!mem->obsolete)
-                       return;
-               mem_cgroup_free(mem);
+               struct mem_cgroup *parent = parent_mem_cgroup(mem);
+               __mem_cgroup_free(mem);
+               if (parent)
+                       mem_cgroup_put(parent);
        }
 }
 
+/*
+ * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
+ */
+static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
+{
+       if (!mem->res.parent)
+               return NULL;
+       return mem_cgroup_from_res_counter(mem->res.parent, res);
+}
 
 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
 static void __init enable_swap_cgroup(void)
@@ -2121,15 +2385,16 @@ static void __init enable_swap_cgroup(void)
 }
 #endif
 
-static struct cgroup_subsys_state *
+static struct cgroup_subsys_state * __ref
 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
 {
        struct mem_cgroup *mem, *parent;
+       long error = -ENOMEM;
        int node;
 
        mem = mem_cgroup_alloc();
        if (!mem)
-               return ERR_PTR(-ENOMEM);
+               return ERR_PTR(error);
 
        for_each_node_state(node, N_POSSIBLE)
                if (alloc_mem_cgroup_per_zone_info(mem, node))
@@ -2146,37 +2411,43 @@ mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
        if (parent && parent->use_hierarchy) {
                res_counter_init(&mem->res, &parent->res);
                res_counter_init(&mem->memsw, &parent->memsw);
+               /*
+                * We increment refcnt of the parent to ensure that we can
+                * safely access it on res_counter_charge/uncharge.
+                * This refcnt will be decremented when freeing this
+                * mem_cgroup(see mem_cgroup_put).
+                */
+               mem_cgroup_get(parent);
        } else {
                res_counter_init(&mem->res, NULL);
                res_counter_init(&mem->memsw, NULL);
        }
-       mem_cgroup_set_inactive_ratio(mem);
-       mem->last_scanned_child = NULL;
+       mem->last_scanned_child = 0;
        spin_lock_init(&mem->reclaim_param_lock);
 
        if (parent)
                mem->swappiness = get_swappiness(parent);
-
+       atomic_set(&mem->refcnt, 1);
        return &mem->css;
 free_out:
-       for_each_node_state(node, N_POSSIBLE)
-               free_mem_cgroup_per_zone_info(mem, node);
-       mem_cgroup_free(mem);
-       return ERR_PTR(-ENOMEM);
+       __mem_cgroup_free(mem);
+       return ERR_PTR(error);
 }
 
-static void mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
+static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
                                        struct cgroup *cont)
 {
        struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
-       mem->obsolete = 1;
-       mem_cgroup_force_empty(mem, false);
+
+       return mem_cgroup_force_empty(mem, false);
 }
 
 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
                                struct cgroup *cont)
 {
-       mem_cgroup_free(mem_cgroup_from_cont(cont));
+       struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
+
+       mem_cgroup_put(mem);
 }
 
 static int mem_cgroup_populate(struct cgroup_subsys *ss,
@@ -2197,10 +2468,12 @@ static void mem_cgroup_move_task(struct cgroup_subsys *ss,
                                struct cgroup *old_cont,
                                struct task_struct *p)
 {
+       mutex_lock(&memcg_tasklist);
        /*
         * FIXME: It's better to move charges of this process from old
         * memcg to new memcg. But it's just on TODO-List now.
         */
+       mutex_unlock(&memcg_tasklist);
 }
 
 struct cgroup_subsys mem_cgroup_subsys = {
@@ -2212,6 +2485,7 @@ struct cgroup_subsys mem_cgroup_subsys = {
        .populate = mem_cgroup_populate,
        .attach = mem_cgroup_move_task,
        .early_init = 0,
+       .use_id = 1,
 };
 
 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP