add support for v4 routes over v6 nexthops #181

rcgoodfellow · 2026-01-08T09:09:48Z

This PR adds support for adding IPv4 routes over IPv6 nexthops. This is needed for BGP unnumbered where nexthops are always IPv6 link local addresses.

There are no new tables added to this PR. Rather IPv6 variants of the exiting IPv4 actions on IPv4 router index table are added.

The API has a version bump to accommodate creating IPv4 routes over IPv6 nexhops as well is listing and fetching IPv4 routes.

This code has been tested in Maghemite CI where we have an mgd daemon feeding unnumbered routes to dendrite and they are installed successfully here.

Nieuwejaar

Could you some tests for this? It's not something that's going to get exercised in a normal racklette or a4x2 run.

Nieuwejaar · 2026-01-13T21:50:50Z

.github/buildomat/jobs/packet-test.sh

 wd=`pwd`
 export WS=$wd
-STARTUP_TIMEOUT=${STARTUP_TIMEOUT:=15}
+STARTUP_TIMEOUT=${STARTUP_TIMEOUT:=45}


I've never seen this be an issue before. Are you seeing this in CI or somewhere else?

I see it regularly when running the tests locally on a development VM.

I've been re-testing this after a rebase. I still hit this on a colo VM. I think we should consider colo VMs first class development environments and supported out of the box. I need to set the timeout to 45 to make things reliable. My colo VM has 8 vCPUs and 32 GB of memory. Perhaps we can come up with a better mechanism than a simple timeout. We could poll dpd until the API comes up with a more pessimistic timeout so the poll will make it work more quickly on faster machines.

At any rate I've put this back at 15 for now and will just have to remember to manually set it on colo VMs

If you're hitting this on a colo VM, then it seems reasonable to bump it up.

Nieuwejaar · 2026-01-13T21:54:24Z

dpd-api/src/lib.rs

        path = "/route/ipv4",
+        versions = ..VERSION_V4_OVER_V6_ROUTES
+    }]
+    async fn route_ipv4_list_v2(


Why is this v2? What was v1?

I've taken a very different approach to route versioning in the PR I sent you, which I think is cleaner - especially with a larger-scale change like I had to make. Basically all the v1 compatibility stuff gets put in a v1 module. I think this will also make it easier to pull out stuff when we decide to drop support for an older version.

I'm not wedded to that implementation style necessarily, but we do need to get our version scheme straight. Right now both PRs represent the first stab at backwards compatibility, and we chose different version numbers.

Sorry - I have too many PRs in my head right now. I meant to point you at the underlay/isolation branch: https://github.com/oxidecomputer/dendrite/blob/isolation/dpd-api/src/lib.rs#L69

I believe I've oriented things in this direction now.

It might be a good idea to have Rain take a look at the versioning stuff. This is going to be dendrite's first implementation of it, and it would be good to be sure we're on the right track.

dpd-client/README-testing.md

rcgoodfellow · 2026-01-14T06:02:22Z

Thanks @Nieuwejaar. I've added a packet test. That packet test surfaced some issues that I fixed.

This code will probably be on the default path for a4x2 as the datacenter topology will probably move to BGP unnumbered.

Nieuwejaar · 2026-01-20T13:48:57Z

dpd/p4/metadata.p4

 	bool nat_ingress_hit;		// NATed packet from uplink -> guest
 	bool nat_ingress_port;		// This port accepts only NAT traffic
 	bool encap_needed;
+	bool resolve_nexthop;		// signals nexthop needs to be resolved


Is it possible to use a variable local to the control rather than adding another field to the meta header?

Not strictly locally, as this signal transcends controls. We could plumb an explicit inout parameter to Router4 and Router6. But I'm not sure how much that gets us and is a bit different to how we're currently plumbing things.

This is fine. I only mentioned it because I recently ran into a case where adding a field to meta caused us to blow the stage limit, but making it an inout didn't. That all went away when I took a simpler approach, but it's now given me one more thing to watch out for.

swadm/src/route.rs

needed for BGP unnumbered

Nieuwejaar

Thanks. This looks good. I ticked the Approve box, but I would like it if you could get Rain to give the backwards compatibility stuff a quick look-over.

sunshowers · 2026-01-20T18:42:22Z

dpd-api/src/v1.rs

+// This Source Code Form is subject to the terms of the Mozilla Public
+// License, v. 2.0. If a copy of the MPL was not distributed with this
+// file, You can obtain one at https://mozilla.org/MPL/2.0/
+//
+// Copyright 2026 Oxide Computer Company
+
+use dpd_types::route::Ipv4Route;
+use oxnet::Ipv4Net;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// Represents all mappings of an IPv4 subnet to a its nexthop target(s).
+#[derive(Debug, Clone, Deserialize, Serialize, JsonSchema)]
+pub struct Ipv4Routes {
+    /// Traffic destined for any address within the CIDR block is routed using
+    /// this information.
+    pub cidr: Ipv4Net,
+    /// All RouteTargets associated with this CIDR
+    pub targets: Vec<Ipv4Route>,
+}


Looks good! Note dpd-api hasn't been transitioned over to the RFD 619 scheme yet (didn't want to step on toes given urgency of other work), so this is good. We'll convert this over to the RFD scheme once things have calmed down.

pulls in - oxidecomputer/dendrite#171 - oxidecomputer/dendrite#175 - oxidecomputer/dendrite#179 - oxidecomputer/dendrite#177 - oxidecomputer/dendrite#181 - oxidecomputer/sidecar-lite#133 - oxidecomputer/softnpu#150

rcgoodfellow added this to the 18 milestone Jan 8, 2026

rcgoodfellow mentioned this pull request Jan 8, 2026

BGP unnumbered oxidecomputer/maghemite#598

Open

rcgoodfellow marked this pull request as ready for review January 9, 2026 23:02

rcgoodfellow requested a review from Nieuwejaar January 9, 2026 23:02

Nieuwejaar reviewed Jan 13, 2026

View reviewed changes

rcgoodfellow force-pushed the ry/v4-over-v6-routes branch 4 times, most recently from 8c5e905 to 4647860 Compare January 19, 2026 21:38

askfongjojo assigned rcgoodfellow Jan 19, 2026

Nieuwejaar reviewed Jan 20, 2026

View reviewed changes

rcgoodfellow added 2 commits January 20, 2026 17:57

add support for v4 routes over v6 nexthops

8539452

needed for BGP unnumbered

swadm v4 over v6

b2123c6

rcgoodfellow force-pushed the ry/v4-over-v6-routes branch from 8268e65 to b2123c6 Compare January 20, 2026 17:59

Nieuwejaar approved these changes Jan 20, 2026

View reviewed changes

sunshowers reviewed Jan 20, 2026

View reviewed changes

rcgoodfellow merged commit 846348b into main Jan 20, 2026
5 checks passed

rcgoodfellow deleted the ry/v4-over-v6-routes branch January 20, 2026 21:25

rcgoodfellow mentioned this pull request Jan 20, 2026

bump dendrite to 181 oxidecomputer/omicron#9692

Merged

add support for v4 routes over v6 nexthops #181

add support for v4 routes over v6 nexthops #181

Uh oh!

Conversation

rcgoodfellow commented Jan 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Nieuwejaar left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

rcgoodfellow commented Jan 14, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Nieuwejaar left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

rcgoodfellow commented Jan 8, 2026 •

edited

Loading