Thread (3 messages) 3 messages, 2 authors, 2d ago

Re: [PATCH net] ipv4: require matching source address for route hint reuse

From: Ido Schimmel <idosch@nvidia.com>
Date: 2026-07-20 12:16:48
Also in: lkml, stable
Subsystem: networking [general], networking [ipv4/ipv6], the rest · Maintainers: "David S. Miller", Eric Dumazet, Jakub Kicinski, Paolo Abeni, David Ahern, Ido Schimmel, Linus Torvalds

On Tue, Jul 14, 2026 at 08:26:17PM +0800, Yizhou Zhao wrote:
IPv4 list receive can reuse a route from the previous skb in the same
receive batch. The current eligibility check only compares the destination
address and TOS before calling ip_route_use_hint().

For forwarded routes, ip_route_use_hint() skips fib_validate_source()
unless the hinted route is local. This means a packet with a different
source address can reuse a forwarding dst created for an earlier packet
and avoid source validation such as strict rp_filter.
I'm not sure why we are skipping source validation for non-local routes.

The comment above ip_route_use_hint() says "Implements all the
saddr-related checks as ip_route_input_slow()". I agree that
ip_route_input_slow() only does source validation for RTN_LOCAL, but for
RTN_UNICAST it is calling ip_mkroute_input(), which eventually calls
fib_validate_source().

Paolo, WDYT about always performing source validation [1]?
In a KASAN QEMU router with strict rp_filter on the ingress device, a 
Why mention KASAN? How is it related to this bug / patch?
bad-only burst was dropped entirely, however, a paired valid/bad burst 
with the same destination/TOS made all of the bad packets pass rp_filter.

Require the source address to match before reusing the hint. Packets from
the same source/destination/TOS still take the fast path; packets whose
source changes go through the normal route lookup and source validation
path.
I agree that it fixes the problem, but we will always pay the
performance penalty, even when rp_filter is disabled. According to
commit 02b24941619f ("ipv4: use dst hint for ipv4 list receive"), there
is still a performance gain when we perform the source validation
per-packet.

[1]
diff --git a/net/ipv4/route.c b/net/ipv4/route.c
index 3f3de5164d6e..89338111793b 100644
--- a/net/ipv4/route.c
+++ b/net/ipv4/route.c
@@ -2194,6 +2194,7 @@ ip_route_use_hint(struct sk_buff *skb, __be32 daddr, __be32 saddr,
 	struct rtable *rt = skb_rtable(hint);
 	struct net *net = dev_net(dev);
 	u32 tag = 0;
+	int oif = 0;
 
 	if (!in_dev)
 		return reason;
@@ -2214,14 +2215,13 @@ ip_route_use_hint(struct sk_buff *skb, __be32 daddr, __be32 saddr,
 	}
 
 	if (!(rt->rt_flags & RTCF_LOCAL))
-		goto skip_validate_source;
+		oif = dst_dev_rcu(&rt->dst)->ifindex;
 
-	reason = fib_validate_source_reason(skb, saddr, daddr, dscp, 0, dev,
+	reason = fib_validate_source_reason(skb, saddr, daddr, dscp, oif, dev,
 					    in_dev, &tag);
 	if (reason)
 		goto martian_source;
 
-skip_validate_source:
 	skb_dst_copy(skb, hint);
 	return SKB_NOT_DROPPED_YET;
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help