David> We see so many issues with ipv6 symbols dependencies in
    David> various subsystems (netfilter, xfrm, etc.).  It is a sign
    David> that we need some kind of long range plan to deal with this
    David> problem.  Since the ipv6 module can't be unloaded anyways,
    David> and it's been broken like that forever, it might make sense
    David> to make ipv6 only available non-modular.  I know people
    David> would dislike this, but the current situation isn't good
    David> either.

Yes, coincidentally I just ran into this problem.  In the context of
reviewing a patch (IPoIB connected mode, which has to deal with
getting too-big packets for a given path) that basically wants to do

#ifdef CONFIG_IPV6
        if (...)
                icmpv6_send(...TOOBIG...)
#endif

I noticed that this exact problem already appears in
net/ipv4/ip_gre.c, which has exactly the same type of code:

#ifdef CONFIG_IPV6
        else if (skb->protocol == htons(ETH_P_IPV6)) {
                struct rt6_info *rt6 = (struct rt6_info*)skb->dst;

                if (rt6 && mtu < dst_mtu(skb->dst) && mtu >= IPV6_MIN_MTU) {
                        if ((tunnel->parms.iph.daddr && 
!MULTICAST(tunnel->parms.iph.daddr)) ||
                            rt6->rt6i_dst.plen == 128) {
                                rt6->rt6i_flags |= RTF_MODIFIED;
                                skb->dst->metrics[RTAX_MTU-1] = mtu;
                        }
                }

                if (mtu >= IPV6_MIN_MTU && mtu < skb->len - tunnel->hlen + 
gre_hlen) {
                        icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu, dev);
                        ip_rt_put(rt);
                        goto tx_error;
                }
        }
#endif

now obviously this means that if ipv6 is built modular, then the
correct handling for too-big packets will never be used.  But on the
other hand, if we convert the test to

#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)

then building with CONFIG_NET_IPGRE=y and CONFIG_IPV6=m will break,
because icmpv6_send() is no longer built in.  And obviously making
NET_IPGRE depend on IPV6 doesn't make sense.

So I hope we can come up with a short-range plan to deal with the
possibility of built-in code calling icmpv6_send() at least...  As you
said, should we just convert IPV6 to a bool instead of a tristate?

Thanks,
  Roland
-
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Reply via email to