Revert "cgroup: Drop task_lock(parent) on cgroup_fork()"
authorTejun Heo <tj@kernel.org>
Fri, 19 Oct 2012 00:52:07 +0000 (17:52 -0700)
committerTejun Heo <tj@kernel.org>
Fri, 19 Oct 2012 21:08:49 +0000 (14:08 -0700)
This reverts commit 7e381b0eb1e1a9805c37335562e8dc02e7d7848c.

The commit incorrectly assumed that fork path always performed
threadgroup_change_begin/end() and depended on that for
synchronization against task exit and cgroup migration paths instead
of explicitly grabbing task_lock().

threadgroup_change is not locked when forking a new process (as
opposed to a new thread in the same process) and even if it were it
wouldn't be effective as different processes use different threadgroup
locks.

Revert the incorrect optimization.

Signed-off-by: Tejun Heo <tj@kernel.org>
LKML-Reference: <20121008020000.GB2575@localhost>
Acked-by: Li Zefan <lizefan@huawei.com>
Bitterly-Acked-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: stable@vger.kernel.org
kernel/cgroup.c

index d1739fc7eb9410435cf40c3334ceb45a10639ec0..75aec12c78a0aeb7c96f761572e5d88777382e95 100644 (file)
@@ -4814,31 +4814,20 @@ static const struct file_operations proc_cgroupstats_operations = {
  *
  * A pointer to the shared css_set was automatically copied in
  * fork.c by dup_task_struct().  However, we ignore that copy, since
- * it was not made under the protection of RCU, cgroup_mutex or
- * threadgroup_change_begin(), so it might no longer be a valid
- * cgroup pointer.  cgroup_attach_task() might have already changed
- * current->cgroups, allowing the previously referenced cgroup
- * group to be removed and freed.
- *
- * Outside the pointer validity we also need to process the css_set
- * inheritance between threadgoup_change_begin() and
- * threadgoup_change_end(), this way there is no leak in any process
- * wide migration performed by cgroup_attach_proc() that could otherwise
- * miss a thread because it is too early or too late in the fork stage.
+ * it was not made under the protection of RCU or cgroup_mutex, so
+ * might no longer be a valid cgroup pointer.  cgroup_attach_task() might
+ * have already changed current->cgroups, allowing the previously
+ * referenced cgroup group to be removed and freed.
  *
  * At the point that cgroup_fork() is called, 'current' is the parent
  * task, and the passed argument 'child' points to the child task.
  */
 void cgroup_fork(struct task_struct *child)
 {
-       /*
-        * We don't need to task_lock() current because current->cgroups
-        * can't be changed concurrently here. The parent obviously hasn't
-        * exited and called cgroup_exit(), and we are synchronized against
-        * cgroup migration through threadgroup_change_begin().
-        */
+       task_lock(current);
        child->cgroups = current->cgroups;
        get_css_set(child->cgroups);
+       task_unlock(current);
        INIT_LIST_HEAD(&child->cg_list);
 }