[patch] xfrm_policy destructor fix

[Ingo Molnar <mingo@elte.hu>]

the patch below fixes a bug that i encountered while running a 
PREEMPT_RT kernel, but i believe it should be fixed in the generic 
kernel too. xfrm_policy_kill() queues a destroyed policy structure to 
the GC list, and unlocks the policy->lock spinlock _after_ that point.  
This created a scenario where GC processing got to the new structure 
first, and kfree()d it - then the write_unlock_bh() was done on the 
already kfreed structure. There is no guarantee that GC processing will 
be done after policy->lock has been dropped and softirq processing has 
been enabled.

Signed-off-by: Ingo Molnar <mingo@elte.hu>

--- linux/net/xfrm/xfrm_policy.c.orig
+++ linux/net/xfrm/xfrm_policy.c
@@ -301,18 +301,22 @@ static void xfrm_policy_gc_task(void *da
 static void xfrm_policy_kill(struct xfrm_policy *policy)
 {
 	write_lock_bh(&policy->lock);
-	if (policy->dead)
-		goto out;
-
+	if (policy->dead) {
+		write_unlock_bh(&policy->lock);
+		return;
+	}
 	policy->dead = 1;
 
 	spin_lock(&xfrm_policy_gc_lock);
 	list_add(&policy->list, &xfrm_policy_gc_list);
+	/*
+	 * Unlock the policy (out of order unlocking), to make sure
+	 * the GC context does not free it with an active lock:
+	 */
+	write_unlock_bh(&policy->lock);
 	spin_unlock(&xfrm_policy_gc_lock);
-	schedule_work(&xfrm_policy_gc_work);
 
-out:
-	write_unlock_bh(&policy->lock);
+	schedule_work(&xfrm_policy_gc_work);
 }
 
 /* Generate new index... KAME seems to generate them ordered by cost