I think the bug is because when creating exceptions, src_addr is not
always set even though fib6_info is in the subtree. (because of
rt6_is_gw_or_nonexthop() check)
However, when looking up for exceptions, we always set src_addr to the
passed in flow->src_addr if fib6_info is in the subtree. That causes
the exception lookup to fail.
I will make it consistent.
However, I don't quite understand the following logic in ip6_rt_cache_alloc():
if (!rt6_is_gw_or_nonexthop(ort)) {
if (ort->fib6_dst.plen != 128 &&
ipv6_addr_equal(&ort->fib6_dst.addr, daddr))
rt->rt6i_flags |= RTF_ANYCAST;
#ifdef CONFIG_IPV6_SUBTREES
if (rt->rt6i_src.plen && saddr) {
rt->rt6i_src.addr = *saddr;
rt->rt6i_src.plen = 128;
}
#endif
}
Why do we need to check that the route is not gateway and has next hop
for updating rt6i_src? I checked the git history and it seems this
part was there from very early on (with some refactor in between)...
From: Stefano Brivio <sbri...@redhat.com>
Date: Tue, May 14, 2019 at 7:33 AM
To: Mikael Magnusson
Cc: Wei Wang, David Ahern, Linux Kernel Network Developers, Martin KaFai Lau
> On Mon, 13 May 2019 23:12:31 -0700
> Wei Wang <wei...@google.com> wrote:
>
> > Thanks Mikael for reporting this issue. And thanks David for the bisection.
> > Let me spend some time to reproduce it and see what is going on.
>
> Mikael, by the way, once this is sorted out, it would be nice if you
> could add your test as a case in tools/testing/selftests/net/pmtu.sh --
> you could probably reuse all the setup parts that are already
> implemented there.
>
> --
> Stefano
