On Tue, Jul 19, 2022 at 09:05:52PM +0200, Stefano Brivio wrote:
> On Tue, 19 Jul 2022 16:20:45 +1000
> David Gibson <david(a)gibson.dropbear.id.au> wrote:
> 
> > On Fri, Jul 15, 2022 at 03:21:40PM +1000, David Gibson wrote:
> > > By default, passt itself attaches to the first host interface with a
> > > default route.  However, when determining the host interface name the tests
> > > implicitly select the *last* host interface: they use a jq expression which
> > > will list all interfaces with default routes, but the way output detection
> > > works in the scripts, it will only pick up the last line.
> > > 
> > > If there are multiple interfaces with default routes on the host, and they
> > > each have a different address, this can cause spurious test
> > > failures.  
> > 
> > It seems this change is not enough to always fix the tests when there
> > are multiple default routes.  I'm still sometimes getting failures,
> > now because passt itself doesn't seem to be picking the interface
> > with the first default route.
> > 
> > I'm wondering if this is because ip(8) is sorting the output, not just
> > presenting it in the same order that the underlying netlink interface
> > does.
> 
> I don't see that happening at least in my environment (and I also can't
> see any code that would sort it, that pretty much comes from
> rtnl_dump_filter_l() in lib/libnetlink.c).
> 
> The netlink "filter", though, is slightly different. For IPv4:
> 
> $ strace -e sendto ip route show >/dev/null
> sendto(3, [{{len=36, type=RTM_GETROUTE, flags=NLM_F_REQUEST|NLM_F_DUMP, seq=1658257027, pid=0}, {rtm_family=AF_INET, rtm_dst_len=0, rtm_src_len=0, rtm_tos=0, rtm_table=RT_TABLE_UNSPEC, rtm_protocol=RTPROT_UNSPEC, rtm_scope=RT_SCOPE_UNIVERSE, rtm_type=RTN_UNSPEC, rtm_flags=0}, {{nla_len=8, nla_type=RTA_TABLE}, RT_TABLE_MAIN}}, {len=0, type=0 /* NLMSG_??? */, flags=0, seq=0, pid=0}], 156, 0, NULL, 0) = 156
> 
> $ strace -e sendto ./passt -f
> sendto(5, {{len=28, type=RTM_GETROUTE, flags=NLM_F_REQUEST|NLM_F_DUMP, seq=0, pid=0}, {rtm_family=AF_INET, rtm_dst_len=0, rtm_src_len=0, rtm_tos=0, rtm_table=RT_TABLE_MAIN, rtm_protocol=RTPROT_UNSPEC, rtm_scope=RT_SCOPE_UNIVERSE, rtm_type=RTN_UNICAST, rtm_flags=0}}, 28, 0, NULL, 0) = 28
> [...]
> 
> and, while I don't think the FIB trie is actually descended in a
> different way, we might still have a slightly different result from the
> kernel (if I recall correctly, I didn't check right now).
> 
> But letting that aside for a moment: if you have two default routes, I
> suppose they have different metrics. If not, what's the intended usage?

Yes, there are two different metrics.  I was thinking after I sent
this that we should sort by metric.

> If yes, we should probably implement a sorting logic in passt, so that
> the route with the lowest metric is picked, and then adjust the jq
> expression to also pick that one.

Agreed.  I'll see if I can implement this.  Up to you whether you drop
this patch or I just do another update on top of it.

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson