xps: must clear sender_cpu before forwarding

John reported that my previous commit added a regression
on his router.

This is because sender_cpu & napi_id share a common location,
so get_xps_queue() can see garbage and perform an out of bound access.

We need to make sure sender_cpu is cleared before doing the transmit,
otherwise any NIC busy poll enabled (skb_mark_napi_id()) can trigger
this bug.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: John <jw@nuclearfallout.net>
Bisected-by: John <jw@nuclearfallout.net>
Fixes: 2bd8248 ("xps: fix xps for stacked devices")
Signed-off-by: David S. Miller <davem@davemloft.net>

include/linux/skbuff.h

@@ -948,6 +948,13 @@ static inline void skb_copy_hash(struct sk_buff *to, const struct sk_buff *from)
 	to->l4_hash = from->l4_hash;
 };
 
+static inline void skb_sender_cpu_clear(struct sk_buff *skb)
+{
+#ifdef CONFIG_XPS
+	skb->sender_cpu = 0;
+#endif
+}
+
 #ifdef NET_SKBUFF_DATA_USES_OFFSET
 static inline unsigned char *skb_end_pointer(const struct sk_buff *skb)
 {

net/core/skbuff.c

@@ -4173,7 +4173,7 @@ void skb_scrub_packet(struct sk_buff *skb, bool xnet)
 	skb->ignore_df = 0;
 	skb_dst_drop(skb);
 	skb->mark = 0;
-	skb->sender_cpu = 0;
+	skb_sender_cpu_clear(skb);
 	skb_init_secmark(skb);
 	secpath_reset(skb);
 	nf_reset(skb);

net/ipv4/ip_forward.c

@@ -67,6 +67,7 @@ static int ip_forward_finish(struct sk_buff *skb)
 	if (unlikely(opt->optlen))
 		ip_forward_options(skb);
 
+	skb_sender_cpu_clear(skb);
 	return dst_output(skb);
 }
 

net/ipv6/ip6_output.c

@@ -318,6 +318,7 @@ static int ip6_forward_proxy_check(struct sk_buff *skb)
 
 static inline int ip6_forward_finish(struct sk_buff *skb)
 {
+	skb_sender_cpu_clear(skb);
 	return dst_output(skb);
 }