Re: VRF destination unreachable

[Stephen Suryaputra]

Thanks David for confirming that this is unintended.

Given that, I think the fix isn't complete because when forwarding
isn't enabled on the original incoming netdev, the ICMP shouldn't be
generated. This diff fixes that case:

diff --git a/net/ipv4/route.c b/net/ipv4/route.c
index a4f44d8..dc40a94 100644
--- a/net/ipv4/route.c
+++ b/net/ipv4/route.c
@@ -937,13 +937,20 @@ static int ip_error(struct sk_buff *skb)
        struct net *net;
        bool send;
        int code;
+       int in_dev_forward;

        /* IP on this device is disabled. */
        if (!in_dev)
                goto out;

        net = dev_net(rt->dst.dev);
-       if (!IN_DEV_FORWARD(in_dev)) {
+       if (netif_is_l3_master(skb->dev)) {
+               in_dev_forward = IN_DEV_FORWARD(
+                       __in_dev_get_rcu(__dev_get_by_index(net,
IPCB(skb)->iif)));
+       } else {
+               in_dev_forward = IN_DEV_FORWARD(in_dev);
+       }
+       if (!in_dev_forward) {
                switch (rt->dst.error) {
                case EHOSTUNREACH:
                        __IP_INC_STATS(net, IPSTATS_MIB_INADDRERRORS);

I can put a formal patch if you like.

Regards,

Stephen.

On Fri, Feb 23, 2018 at 3:58 PM, David Ahern <dsah...@gmail.com> wrote:
> On 2/23/18 10:49 AM, Stephen Suryaputra wrote:
>> Greetings,
>>
>> We found that ICMP destination unreachable isn't sent if VRF
>> forwarding isn't configured, i.e.
>> /proc/sys/net/ipv4/conf/<vrf_net_device>/forwarding isn't set. The
>> relevant code is:
>>
>> static int ip_error(struct sk_buff *skb)
>> {
>> ...
>>         // in_dev is the vrf net_device
>>         if (!IN_DEV_FORWARD(in_dev)) {
>>                 switch (rt->dst.error) {
>>                 case EHOSTUNREACH:
>>                         __IP_INC_STATS(net, IPSTATS_MIB_INADDRERRORS);
>>                         break;
>>
>>                 case ENETUNREACH:
>>                         __IP_INC_STATS(net, IPSTATS_MIB_INNOROUTES);
>>                         break;
>>                 }
>>                 goto out;
>>         }
>> ...
>> out:    kfree_skb(skb);
>>         return 0;
>> }
>>
>> The question: is it intended to be set? The basic forwarding seems to
>> be working without. We do set it on the slave net devices.
>
> Unintended side effect of VRF as a netdev. This should fix it:
>
> diff --git a/net/ipv4/route.c b/net/ipv4/route.c
> index 5ca7415cd48c..d59d005fb7c5 100644
> --- a/net/ipv4/route.c
> +++ b/net/ipv4/route.c
> @@ -944,7 +944,7 @@ static int ip_error(struct sk_buff *skb)
>                 goto out;
>
>         net = dev_net(rt->dst.dev);
> -       if (!IN_DEV_FORWARD(in_dev)) {
> +       if (!IN_DEV_FORWARD(in_dev) && !netif_is_l3_master(skb->dev)) {
>                 switch (rt->dst.error) {
>                 case EHOSTUNREACH:
>                         __IP_INC_STATS(net, IPSTATS_MIB_INADDRERRORS);