cgroups: fix a race between rmdir and remount

When a cgroup is removed, it's unlinked from its parent's children list,
but not actually freed until the last dentry on it is released (at which
point cgrp->root->number_of_cgroups is decremented).

Currently rebind_subsystems checks for the top cgroup's child list being
empty in order to rebind subsystems into or out of a hierarchy - this can
result in the set of subsystems bound to a hierarchy being
removed-but-not-freed cgroup.

The simplest fix for this is to forbid remounts that change the set of
subsystems on a hierarchy that has removed-but-not-freed cgroups.  This
bug can be reproduced via:

mkdir /mnt/cg
mount -t cgroup -o ns,freezer cgroup /mnt/cg
mkdir /mnt/cg/foo
sleep 1h < /mnt/cg/foo &
rmdir /mnt/cg/foo
mount -t cgroup -o remount,ns,devices,freezer cgroup /mnt/cg
kill $!

Though the above will cause oops in -mm only but not mainline, but the bug
can cause memory leak in mainline (and even oops)

Signed-off-by: Paul Menage <menage@google.com>
Reviewed-by: Li Zefan <lizf@cn.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>