Re: PMTU discovery broken in Linux for UDP/raw application if the socket is not bound to a device

[ adding netdev so others know ]

On 10/9/18 3:38 AM, Preethi Ramachandra wrote:

Hi David,

I tested your fix, Linux is updating PMTU successfully. 

ok, I'll send a formal patch

Thanks,
Preethi

On 10/7/18, 8:59 AM, "David Ahern" [off-list ref] wrote:

    The correct mailing list is netdev@vger.kernel.org (added)
    
    non-text emails will be rejected.
    
    
    On 10/3/18 10:15 PM, Preethi Ramachandra wrote:
    > Hi,
    > 
    >  
    > 
    > While testing the PMTU discovery for UDP/raw applications, Linux is not
    > doing PMTU discovery if the UDP server socket is not bound to a device.
    >  In the scenario we are testing there could be multiple VRF devices
    > created and an application like UDP/RAW can use a common socket for all
    > vrf devices. While sending packet IP_PKTINFO socket option can be used
    > to specify the vrf interface through which packet will be sent out. In
    > this case, when packet too big icmp6 error message comes back to Linux
    > on a vrf device, a route lookup is done on default routing-table(0) for
    > src/dst address which case, the route will not be found and packet is
    > dropped. If the route lookup happened with proper VRF device (packet’s
    > incoming index), the route lookup succeeds,  PMTU discovery is successful.
    > 
    >  
    > 
    > This might need a fix, please take a look.
    > 
    >  
    > 
    > *Linux version *
    > 
    >  
    > 
    > Linux  4.8.24
    > 
    >  
    > 
    > *Code flow *
    > 
    >  
    > 
    > Linux code where it expects socket’s bound device in order for PMTU
    > discovery to happen.
    > 
    > *void ip6_sk_update_pmtu*(struct sk_buff *skb, struct sock *sk, __be32 mtu)
    > 
    > {
    > 
    >                     struct dst_entry *dst;
    > 
    >  
    > 
    >                     ip6_update_pmtu(skb, sock_net(sk), mtu,
    > 
    >                                                             
    > sk->sk_bound_dev_if, sk->sk_mark, sk->sk_uid);*<<<<< This is the point
    > where it expects socket’s sk_bound_dev_if to be set. In our testing this
    > is actually 0, since the socket is not really bound to a vrf device.*
    
    Try this based on top of tree for 4.19-next (whitespace damaged on paste
    so you'll need to manually apply and handle differences with 4.8):
    
    diff --git a/net/ipv6/route.c b/net/ipv6/route.c
    index 6c1d817151ca..50b95b48b911 100644
    --- a/net/ipv6/route.c
    +++ b/net/ipv6/route.c
    @@ -2360,10 +2360,13 @@ EXPORT_SYMBOL_GPL(ip6_update_pmtu);
    
     void ip6_sk_update_pmtu(struct sk_buff *skb, struct sock *sk, __be32 mtu)
     {
    +       int oif = sk->sk_bound_dev_if;
            struct dst_entry *dst;
    
    -       ip6_update_pmtu(skb, sock_net(sk), mtu,
    -                       sk->sk_bound_dev_if, sk->sk_mark, sk->sk_uid);
    +       if (!oif && skb->dev)
    +               oif = l3mdev_master_ifindex(skb->dev);
    +
    +       ip6_update_pmtu(skb, sock_net(sk), mtu, oif, sk->sk_mark,
    sk->sk_uid);
    
            dst = __sk_dst_get(sk);
            if (!dst || !dst->obsolete ||