Re: [PATCH 2/3] ipv4/icmp: l3mdev: Perform icmp error route lookup on source device routing table

[David Ahern]

On 8/11/20 1:50 PM, Mathieu Desnoyers wrote:
> As per RFC792, ICMP errors should be sent to the source host.
> 
> However, in configurations with Virtual Routing and Forwarding tables,
> looking up which routing table to use is currently done by using the
> destination net_device.
> 
> commit 9d1a6c4ea43e ("net: icmp_route_lookup should use rt dev to
> determine L3 domain") changes the interface passed to
> l3mdev_master_ifindex() and inet_addr_type_dev_table() from skb_in->dev
> to skb_dst(skb_in)->dev. This effectively uses the destination device
> rather than the source device for choosing which routing table should be
> used to lookup where to send the ICMP error.
> 
> Therefore, if the source and destination interfaces are within separate
> VRFs, or one in the global routing table and the other in a VRF, looking
> up the source host in the destination interface's routing table will
> fail if the destination interface's routing table contains no route to
> the source host.
> 
> One observable effect of this issue is that traceroute does not work in
> the following cases:
> 
> - Route leaking between global routing table and VRF
> - Route leaking between VRFs
> 
> Preferably use the source device routing table when sending ICMP error
> messages. If no source device is set, fall-back on the destination
> device routing table.
> 
> Fixes: 9d1a6c4ea43e ("net: icmp_route_lookup should use rt dev to determine L3 domain")
> Link: https://tools.ietf.org/html/rfc792
> Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx>
> Cc: David Ahern <dsahern@xxxxxxxxxx>
> Cc: David S. Miller <davem@xxxxxxxxxxxxx>
> Cc: netdev@xxxxxxxxxxxxxxx
> ---
>  net/ipv4/icmp.c | 15 +++++++++++++--
>  1 file changed, 13 insertions(+), 2 deletions(-)
> 
> diff --git a/net/ipv4/icmp.c b/net/ipv4/icmp.c
> index cf36f955bfe6..1eb83d82ec68 100644
> --- a/net/ipv4/icmp.c
> +++ b/net/ipv4/icmp.c
> @@ -465,6 +465,7 @@ static struct rtable *icmp_route_lookup(struct net *net,
>  					int type, int code,
>  					struct icmp_bxm *param)
>  {
> +	struct net_device *route_lookup_dev = NULL;
>  	struct rtable *rt, *rt2;
>  	struct flowi4 fl4_dec;
>  	int err;
> @@ -479,7 +480,17 @@ static struct rtable *icmp_route_lookup(struct net *net,
>  	fl4->flowi4_proto = IPPROTO_ICMP;
>  	fl4->fl4_icmp_type = type;
>  	fl4->fl4_icmp_code = code;
> -	fl4->flowi4_oif = l3mdev_master_ifindex(skb_dst(skb_in)->dev);
> +	/*
> +	 * The device used for looking up which routing table to use is
> +	 * preferably the source whenever it is set, which should ensure
> +	 * the icmp error can be sent to the source host, else fallback
> +	 * on the destination device.
> +	 */
> +	if (skb_in->dev)
> +		route_lookup_dev = skb_in->dev;
> +	else if (skb_dst(skb_in))
> +		route_lookup_dev = skb_dst(skb_in)->dev;
> +	fl4->flowi4_oif = l3mdev_master_ifindex(route_lookup_dev);
>  
>  	security_skb_classify_flow(skb_in, flowi4_to_flowi(fl4));
>  	rt = ip_route_output_key_hash(net, fl4, skb_in);
> @@ -503,7 +514,7 @@ static struct rtable *icmp_route_lookup(struct net *net,
>  	if (err)
>  		goto relookup_failed;
>  
> -	if (inet_addr_type_dev_table(net, skb_dst(skb_in)->dev,
> +	if (inet_addr_type_dev_table(net, route_lookup_dev,
>  				     fl4_dec.saddr) == RTN_LOCAL) {
>  		rt2 = __ip_route_output_key(net, &fl4_dec);
>  		if (IS_ERR(rt2))
> 

ICMP's can be generated in many locations:
1. forward path - I think the skb_in dev is always set,

2. ingress and upper layer protocols -  dev is dropped prior to
transport layers, so, for example, UDP sending port unreachable calls
icmp_send with skb_in->dev set to NULL.

3. local packets and egress - e.g., link failures and here I believe skb
dev is set.

If in and out are in the same L3 domain, either device works where for
VRF route leaking with the forward path in and out are in separate
domains so yes you want the ingress device.

This change seems fine to me and I have not seen any issues with
existing selftests.

Reviewed-by: David Ahern <dsahern@xxxxxxxxxx>


But I did notice that unreachable / fragmentation needed messages are
NOT working with this change. You can see that by changing the MTU of
eth1 in r1 to 1400 and running:
   ip netns exec h1 ping -s 1450 -Mdo -c1 172.16.2.2

You really should get that working as well with VRF route leaking.