On Thu, Jul 5, 2018 at 2:36 AM, David Ahern <dsah...@gmail.com> wrote:
> On 7/4/18 11:56 AM, Xin Long wrote:
>
>>> your commands are not a proper test. The test should succeed and fail
>>> based on the routing lookup, not iptables rules.
>> A proper test can be done easily with netns, as vrf can't isolate much.
>> I don't want to bother forwarding/ directory with netns, so I will probably
>> just drop this selftest, and let the feature patch go first.
>>
>
> BTW, VRF isolates at the routing layer and this is a routing change. We
> need to understand why it does not work with VRF. Perhaps another tweak
> is needed for VRF.
One problem was that the peer may not use the address on the dev
that echo_request comes from as the src IP of echo_reply when the
echo_request's dst IP is broadcast, but try to get another one by
looking up a route without ".flowi4_oif" set. See:
icmp_reply()->fib_compute_spec_dst():
struct flowi4 fl4 = {
.flowi4_iif = LOOPBACK_IFINDEX,
.daddr = ip_hdr(skb)->saddr,
.flowi4_tos = RT_TOS(ip_hdr(skb)->tos),
.flowi4_scope = scope,
.flowi4_mark = IN_DEV_SRC_VMARK(in_dev) ? skb->mark : 0,
};
if (!fib_lookup(net, &fl4, &res, 0))
return FIB_RES_PREFSRC(net, res);
Without ".flowi4_oif" set, it won't match the vrf route. That's why
I had to make h2 NOT into a vrf so that h1 can get the echo_reply.
But it can't tell if this echo_reply is from h2 or r1, as r1's echo_reply
will also use the same src IP which is actually got from main route
space as ".flowi4_oif" is not set.
(hope I this description is clear to you) :)
So i'm not sure if we can do any tweak for VRF.
