netfilter: nf_dup_netdev: do not push mac header a second time

[ Upstream commit 574a5b8 ]

Eric reports skb_under_panic when using dup/fwd via bond+egress hook.
Before pushing mac header, we should make sure that we're called from
ingress to put back what was pulled earlier.

In egress case, the MAC header is already there; we should leave skb
alone.

While at it be more careful here: skb might have been altered and
headroom reduced, so add a skb_cow() before so that headroom is
increased if necessary.

nf_do_netdev_egress() assumes skb ownership (it normally ends with
a call to dev_queue_xmit), so we must free the packet on error.

Fixes: f87b946 ("netfilter: nft_fwd_netdev: Support egress hook")
Reported-by: Eric Garver <eric@garver.life>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>

net/netfilter/nf_dup_netdev.c

@@ -13,10 +13,16 @@
 #include <net/netfilter/nf_tables_offload.h>
 #include <net/netfilter/nf_dup_netdev.h>
 
-static void nf_do_netdev_egress(struct sk_buff *skb, struct net_device *dev)
+static void nf_do_netdev_egress(struct sk_buff *skb, struct net_device *dev,
+				enum nf_dev_hooks hook)
 {
-	if (skb_mac_header_was_set(skb))
+	if (hook == NF_NETDEV_INGRESS && skb_mac_header_was_set(skb)) {
+		if (skb_cow_head(skb, skb->mac_len)) {
+			kfree_skb(skb);
+			return;
+		}
 		skb_push(skb, skb->mac_len);
+	}
 
 	skb->dev = dev;
 	skb_clear_tstamp(skb);
@@ -33,7 +39,7 @@ void nf_fwd_netdev_egress(const struct nft_pktinfo *pkt, int oif)
 		return;
 	}
 
-	nf_do_netdev_egress(pkt->skb, dev);
+	nf_do_netdev_egress(pkt->skb, dev, nft_hook(pkt));
 }
 EXPORT_SYMBOL_GPL(nf_fwd_netdev_egress);
 
@@ -48,7 +54,7 @@ void nf_dup_netdev_egress(const struct nft_pktinfo *pkt, int oif)
 
 	skb = skb_clone(pkt->skb, GFP_ATOMIC);
 	if (skb)
-		nf_do_netdev_egress(skb, dev);
+		nf_do_netdev_egress(skb, dev, nft_hook(pkt));
 }
 EXPORT_SYMBOL_GPL(nf_dup_netdev_egress);