memcg: use css_get/put for swap memcg
authorLi Zefan <lizefan@huawei.com>
Mon, 8 Jul 2013 23:00:34 +0000 (16:00 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 9 Jul 2013 17:33:24 +0000 (10:33 -0700)
Use css_get/put instead of mem_cgroup_get/put.  A simple replacement
will do.

The historical reason that memcg has its own refcnt instead of always
using css_get/put, is that cgroup couldn't be removed if there're still
css refs, so css refs can't be used as long-lived reference.  The
situation has changed so that rmdir a cgroup will succeed regardless css
refs, but won't be freed until css refs goes down to 0.

Signed-off-by: Li Zefan <lizefan@huawei.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Glauber Costa <glommer@openvz.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/memcontrol.c

index bdc9582..76c0c99 100644 (file)
@@ -4231,12 +4231,12 @@ __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
        unlock_page_cgroup(pc);
        /*
         * even after unlock, we have memcg->res.usage here and this memcg
-        * will never be freed.
+        * will never be freed, so it's safe to call css_get().
         */
        memcg_check_events(memcg, page);
        if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
                mem_cgroup_swap_statistics(memcg, true);
-               mem_cgroup_get(memcg);
+               css_get(&memcg->css);
        }
        /*
         * Migration does not charge the res_counter for the
@@ -4348,7 +4348,7 @@ mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
 
        /*
         * record memcg information,  if swapout && memcg != NULL,
-        * mem_cgroup_get() was called in uncharge().
+        * css_get() was called in uncharge().
         */
        if (do_swap_account && swapout && memcg)
                swap_cgroup_record(ent, css_id(&memcg->css));
@@ -4379,7 +4379,7 @@ void mem_cgroup_uncharge_swap(swp_entry_t ent)
                if (!mem_cgroup_is_root(memcg))
                        res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
                mem_cgroup_swap_statistics(memcg, false);
-               mem_cgroup_put(memcg);
+               css_put(&memcg->css);
        }
        rcu_read_unlock();
 }
@@ -4413,11 +4413,14 @@ static int mem_cgroup_move_swap_account(swp_entry_t entry,
                 * 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.
+                * improvement. But we cannot postpone css_get(to)  because if
+                * the process that has been moved to @to does swap-in, the
+                * refcount of @to might be decreased to 0.
+                *
+                * We are in attach() phase, so the cgroup is guaranteed to be
+                * alive, so we can just call css_get().
                 */
-               mem_cgroup_get(to);
+               css_get(&to->css);
                return 0;
        }
        return -EINVAL;
@@ -6718,6 +6721,7 @@ static void __mem_cgroup_clear_mc(void)
 {
        struct mem_cgroup *from = mc.from;
        struct mem_cgroup *to = mc.to;
+       int i;
 
        /* we must uncharge all the leftover precharges from mc.to */
        if (mc.precharge) {
@@ -6738,7 +6742,9 @@ static void __mem_cgroup_clear_mc(void)
                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);
+
+               for (i = 0; i < mc.moved_swap; i++)
+                       css_put(&mc.from->css);
 
                if (!mem_cgroup_is_root(mc.to)) {
                        /*
@@ -6748,7 +6754,7 @@ static void __mem_cgroup_clear_mc(void)
                        res_counter_uncharge(&mc.to->res,
                                                PAGE_SIZE * mc.moved_swap);
                }
-               /* we've already done mem_cgroup_get(mc.to) */
+               /* we've already done css_get(mc.to) */
                mc.moved_swap = 0;
        }
        memcg_oom_recover(from);