memcg: improve performance in moving swap charge
authorDaisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Wed, 10 Mar 2010 23:22:18 +0000 (15:22 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 12 Mar 2010 23:52:36 +0000 (15:52 -0800)
Try to reduce overheads in moving swap charge by:

- Adds a new function(__mem_cgroup_put), which takes "count" as a arg and
  decrement mem->refcnt by "count".
- Removed res_counter_uncharge, css_put, and mem_cgroup_put from the path
  of moving swap account, and consolidate all of them into mem_cgroup_clear_mc.
  We cannot do that about mc.to->refcnt.

These changes reduces the overhead from 1.35sec to 0.9sec to move charges
of 1G anonymous memory(including 500MB swap) in my test environment.

Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Cc: Balbir Singh <balbir@linux.vnet.ibm.com>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Cc: Paul Menage <menage@google.com>
Cc: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/memcontrol.c

index e883198baf81ecea1605753a3a9b768487f6693f..b00ec74a4c183c6ef83ba1cc37b2346c8069161e 100644 (file)
@@ -255,6 +255,7 @@ static struct move_charge_struct {
        struct mem_cgroup *to;
        unsigned long precharge;
        unsigned long moved_charge;
+       unsigned long moved_swap;
        struct task_struct *moving_task;        /* a task moving charges */
        wait_queue_head_t waitq;                /* a waitq for other context */
 } mc = {
@@ -2277,6 +2278,7 @@ void mem_cgroup_uncharge_swap(swp_entry_t ent)
  * @entry: swap entry to be moved
  * @from:  mem_cgroup which the entry is moved from
  * @to:  mem_cgroup which the entry is moved to
+ * @need_fixup: whether we should fixup res_counters and refcounts.
  *
  * It succeeds only when the swap_cgroup's record for this entry is the same
  * as the mem_cgroup's id of @from.
@@ -2287,7 +2289,7 @@ void mem_cgroup_uncharge_swap(swp_entry_t ent)
  * both res and memsw, and called css_get().
  */
 static int mem_cgroup_move_swap_account(swp_entry_t entry,
-                               struct mem_cgroup *from, struct mem_cgroup *to)
+               struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
 {
        unsigned short old_id, new_id;
 
@@ -2295,27 +2297,36 @@ static int mem_cgroup_move_swap_account(swp_entry_t entry,
        new_id = css_id(&to->css);
 
        if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
-               if (!mem_cgroup_is_root(from))
-                       res_counter_uncharge(&from->memsw, PAGE_SIZE);
                mem_cgroup_swap_statistics(from, false);
-               mem_cgroup_put(from);
+               mem_cgroup_swap_statistics(to, true);
                /*
-                * we charged both to->res and to->memsw, so we should uncharge
-                * to->res.
+                * This function is only called from task migration context now.
+                * It postpones res_counter and refcount handling till the end
+                * of task migration(mem_cgroup_clear_mc()) for performance
+                * improvement. But we cannot postpone mem_cgroup_get(to)
+                * because if the process that has been moved to @to does
+                * swap-in, the refcount of @to might be decreased to 0.
                 */
-               if (!mem_cgroup_is_root(to))
-                       res_counter_uncharge(&to->res, PAGE_SIZE);
-               mem_cgroup_swap_statistics(to, true);
                mem_cgroup_get(to);
-               css_put(&to->css);
-
+               if (need_fixup) {
+                       if (!mem_cgroup_is_root(from))
+                               res_counter_uncharge(&from->memsw, PAGE_SIZE);
+                       mem_cgroup_put(from);
+                       /*
+                        * we charged both to->res and to->memsw, so we should
+                        * uncharge to->res.
+                        */
+                       if (!mem_cgroup_is_root(to))
+                               res_counter_uncharge(&to->res, PAGE_SIZE);
+                       css_put(&to->css);
+               }
                return 0;
        }
        return -EINVAL;
 }
 #else
 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
-                               struct mem_cgroup *from, struct mem_cgroup *to)
+               struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
 {
        return -EINVAL;
 }
@@ -3398,9 +3409,9 @@ static void mem_cgroup_get(struct mem_cgroup *mem)
        atomic_inc(&mem->refcnt);
 }
 
-static void mem_cgroup_put(struct mem_cgroup *mem)
+static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
 {
-       if (atomic_dec_and_test(&mem->refcnt)) {
+       if (atomic_sub_and_test(count, &mem->refcnt)) {
                struct mem_cgroup *parent = parent_mem_cgroup(mem);
                __mem_cgroup_free(mem);
                if (parent)
@@ -3408,6 +3419,11 @@ static void mem_cgroup_put(struct mem_cgroup *mem)
        }
 }
 
+static void mem_cgroup_put(struct mem_cgroup *mem)
+{
+       __mem_cgroup_put(mem, 1);
+}
+
 /*
  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
  */
@@ -3789,6 +3805,29 @@ static void mem_cgroup_clear_mc(void)
                __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
                mc.moved_charge = 0;
        }
+       /* we must fixup refcnts and charges */
+       if (mc.moved_swap) {
+               WARN_ON_ONCE(mc.moved_swap > INT_MAX);
+               /* uncharge swap account from the old cgroup */
+               if (!mem_cgroup_is_root(mc.from))
+                       res_counter_uncharge(&mc.from->memsw,
+                                               PAGE_SIZE * mc.moved_swap);
+               __mem_cgroup_put(mc.from, mc.moved_swap);
+
+               if (!mem_cgroup_is_root(mc.to)) {
+                       /*
+                        * we charged both to->res and to->memsw, so we should
+                        * uncharge to->res.
+                        */
+                       res_counter_uncharge(&mc.to->res,
+                                               PAGE_SIZE * mc.moved_swap);
+                       VM_BUG_ON(test_bit(CSS_ROOT, &mc.to->css.flags));
+                       __css_put(&mc.to->css, mc.moved_swap);
+               }
+               /* we've already done mem_cgroup_get(mc.to) */
+
+               mc.moved_swap = 0;
+       }
        mc.from = NULL;
        mc.to = NULL;
        mc.moving_task = NULL;
@@ -3818,11 +3857,13 @@ static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
                        VM_BUG_ON(mc.to);
                        VM_BUG_ON(mc.precharge);
                        VM_BUG_ON(mc.moved_charge);
+                       VM_BUG_ON(mc.moved_swap);
                        VM_BUG_ON(mc.moving_task);
                        mc.from = from;
                        mc.to = mem;
                        mc.precharge = 0;
                        mc.moved_charge = 0;
+                       mc.moved_swap = 0;
                        mc.moving_task = current;
 
                        ret = mem_cgroup_precharge_mc(mm);
@@ -3883,8 +3924,12 @@ put:                     /* is_target_pte_for_mc() gets the page */
                        break;
                case MC_TARGET_SWAP:
                        ent = target.ent;
-                       if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to))
+                       if (!mem_cgroup_move_swap_account(ent,
+                                               mc.from, mc.to, false)) {
                                mc.precharge--;
+                               /* we fixup refcnts and charges later. */
+                               mc.moved_swap++;
+                       }
                        break;
                default:
                        break;