memcg: fix wrong VM_BUG_ON() in try_charge()'s mm->owner check

At __mem_cgroup_try_charge(), VM_BUG_ON(!mm->owner) is checked.
But as commented in mem_cgroup_from_task(), mm->owner can be NULL
in some racy case. This check of VM_BUG_ON() is bad.

A possible story to hit this is at swapoff()->try_to_unuse(). It passes
mm_struct to mem_cgroup_try_charge_swapin() while mm->owner is NULL. If we
can't get proper mem_cgroup from swap_cgroup information, mm->owner is used
as charge target and we see NULL.

Cc: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reported-by: Hugh Dickins <hughd@google.com>
Reported-by: Thomas Meyer <thomas@m3y3r.de>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: Balbir Singh <balbir@linux.vnet.ibm.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: stable@kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 7a22b41..00bb8a6 100644 (file)
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -1925,19 +1925,18 @@ again:
 
                rcu_read_lock();
                p = rcu_dereference(mm->owner);
-               VM_BUG_ON(!p);
                /*
-                * because we don't have task_lock(), "p" can exit while
-                * we're here. In that case, "mem" can point to root
-                * cgroup but never be NULL. (and task_struct itself is freed
-                * by RCU, cgroup itself is RCU safe.) Then, we have small
-                * risk here to get wrong cgroup. But such kind of mis-account
-                * by race always happens because we don't have cgroup_mutex().
-                * It's overkill and we allow that small race, here.
+                * Because we don't have task_lock(), "p" can exit.
+                * In that case, "mem" can point to root or p can be NULL with
+                * race with swapoff. Then, we have small risk of mis-accouning.
+                * But such kind of mis-account by race always happens because
+                * we don't have cgroup_mutex(). It's overkill and we allo that
+                * small race, here.
+                * (*) swapoff at el will charge against mm-struct not against
+                * task-struct. So, mm->owner can be NULL.
                 */
                mem = mem_cgroup_from_task(p);
-               VM_BUG_ON(!mem);
-               if (mem_cgroup_is_root(mem)) {
+               if (!mem || mem_cgroup_is_root(mem)) {
                        rcu_read_unlock();
                        goto done;
                }