routing: Allow nexthops of another family in a nexthop group

[w453y.me_gmail.com]

Summary

An IPv4 route with IPv6 gateways (RFC 5549, used by FRR's BGP
unnumbered) can have only one path: adding the second one fails with
EINVAL. Nexthops with an IPv6 gateway are linked to the IPv6
nh_control by nhop_get_rh(), and nhgrp_alloc() requires all members
to be in the nh_control of the group.

Accept members from the same fib and upper family. Nexthop indexes
are allocated per nh_control, so an IPv4 and an IPv6 member can have
the same index: sort the members by neighbor family and then index,
and detect duplicates by pointer.

decompose_change_notification() walks the old and new group in that
order, so use the same compare function there. Otherwise deleting a
path from a mixed group sends RTM_DELETE with the wrong gateway on
the routing socket.

netstat -On still shows IPv6 members of an IPv4 group wrong. That
needs the family in the group sysctl and is done separately.

Signed-off-by: Abdul Wasey <[e-mail]>

Test Plan

• Repro without the patch: route -4 add -net 198.18.0.0/24 -inet6 fe80::2%epair0a, then the same with fe80::3%epair1a: the second add fails with EINVAL. With the patch the route has both paths, and FRR installs its BGP unnumbered ECMP route and forwards over both.
• New tests in tests/sys/net/routing/test_rtsock_multipath.py. test_rtm_multipath4_mixed_notify fails without the decompose_change_notification() change: assert [(2, '2001:db8::10')] == [(2, '192.0.2.16')].
• kyua sys/net/routing: 85/86 pass (test_ecmp_routes_by_event fails on the unpatched kernel too).
• kyua sys/netlink: 116/117 pass, 1 skipped.

[pouria]

Nice work.
I still haven't tested the code itself yet, and I'm worried about a couple of things.
Let me test and think about those, and I will let you know.

sys/net/route/nhgrp_ctl.c

91–118

wn_cmp_idx is static and nhgrp_wn_cmp is public.
I suggest keeping the wn_cmp_idx as before and implementing nhgrp_wn_cmp the other way around.

558–562

You already explained it below, no need to state the obvious.

sys/net/route/route_helpers.c

354–355

You can remove this comment now.

[w453y.me_gmail.com]

Thanks! Done in this update: wn_cmp_idx() is the compare again and nhgrp_wn_cmp() just calls it, and both comments are gone.