[PATCH] avoid race condition in pick_next_task_fair in kernel/sched_fair.c

[shenghui]

Hi,

      I walked some code in kernel/sched_fair.c in version 2.6.35-rc3, and
got the following potential failure:

static struct task_struct *pick_next_task_fair(struct rq *rq)
{
...
	if (!cfs_rq->nr_running)
		return NULL;

	do {
		se = pick_next_entity(cfs_rq);
		set_next_entity(cfs_rq, se);
		cfs_rq = group_cfs_rq(se);
	} while (cfs_rq);
...
}

/*
 * The dequeue_task method is called after nr_running is
 * decreased. We remove the task from the rbtree and
 * update the fair scheduling stats:
 */
static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
{
...
		dequeue_entity(cfs_rq, se, flags);
...
}

static void
dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
{
...
	if (se != cfs_rq->curr)
		__dequeue_entity(cfs_rq, se);
        account_entity_dequeue(cfs_rq, se);
	update_min_vruntime(cfs_rq);
...
}

dequeue_task_fair() is used to dequeue some task, and it calls
dequeue_entity() to finish the job. While dequeue_entity() calls
__dequeue_entity() first, then calls account_entity_dequeue()
to do substraction on cfs_rq->nr_running.
Here, if one task is dequeued, then cfs_rq->nr_running will be
changed later, e.g 1 to 0.
In the pick_next_task_fair(), the if check will get passed before
nr_running is set 0 and the following while structure is executed.

	do {
		se = pick_next_entity(cfs_rq);
		set_next_entity(cfs_rq, se);
		cfs_rq = group_cfs_rq(se);
	} while (cfs_rq);

We may get se set NULL here, and set_next_entity may deference
NULL pointer, which can lead to Oops.

I think some lock on the metadata can fix this issue, but we may
change plenty of code to add support for lock. I think the easist
way is just substacting nr_running before dequing tasks.

Following is my patch, please check it.