Fixes found while running Multipath TCP across real links that drop and come back.
Open Source
Patches and bug reports I have sent to other people's projects. Most of these came out of hitting something while building or running my own stuff, which is usually the only reason anyone reads someone else's code that closely.
Mostly found by running bfdd with xdp-bfd as its data plane, the rest by adding demand mode and authentication.
- #23465mergedStops bfdd opening echo sockets when a data plane handles echo, which left a peer’s IPv6 echo answered by the wrong side or refused, depending on startup order.Read the write-up
- #23453mergedSends no echo interval to a data plane when the peer says it can’t receive echo.Read the write-up
Show 15 more
- #23495mergedSends IPv6 echo with the same traffic class as the rest of BFD, so a congested switch doesn’t drop it first.Read the write-up
- #23496mergedOpens the echo sockets for a session enabled after bfdd starts. Since FRR 10.7.0 those sessions got no sockets and kept flapping.Read the write-up
- #23463mergedLets a data plane tell bfdd what it supports, so authenticated sessions only go to one that can handle them.Read the write-up
- #23464mergedForgets the peer’s discriminator after a detection time, fixing a case where a restarted peer stayed down.Read the write-up
- #23331mergedSends authentication keys and their lifetimes to the data plane. Before this, an offloaded session with a keychain went out unauthenticated while the CLI said auth was on.Read the write-up
- #23282mergedA keychain with no usable key no longer runs the session unauthenticated. SHA1 keys of 17 to 20 bytes were being rejected and the session quietly fell back to no auth.Read the write-up
- #23281mergedMade keyed SHA1 sequence number checking actually run. It never did, so replayed packets were accepted.Read the write-up
- #23284mergedChecks the packet’s state, not the local session’s, when Your Discriminator is zero. In demand mode the old check left a restarted peer locked out.Read the write-up
- #23023mergedImplemented demand mode (RFC 5880 section 6.6), which bfdd declared but never did anything with.Read the write-up
- #22694mergedFixed "show bfd peers counters" dropping the data plane connection every second run, and added a fake data plane to FRR’s test tree so this path can be tested.Read the write-up
- #22692mergedSession delete messages queued on shutdown are now actually sent, instead of being freed with the buffer.Read the write-up
- #22920mergedIPv6 echo is sent from the session’s own address. The kernel could pick one the peer rejects, so turning on echo took the session down.Read the write-up
- #22805mergedOffloaded sessions now negotiate the echo interval with the peer, instead of sending at whatever was configured locally.Read the write-up
- #22645mergedWith more than 40 to 58 sessions, depending on config, the extra sessions were silently lost when the data plane registration burst overflowed an 8KB buffer. They now register, or fall back to software BFD if they can’t.Read the write-up
- #22621mergedFixed the UNIX socket data plane transport, which failed every connect because bfdd passed the size of an address union instead of the real address length.Read the write-up
A BGP implementation in Go, also used as a library by Kubernetes networking projects. These fix how its BFD support treats a session. I filed the bugs first, then fixed them.
- #3781mergedWhen the peer turned its BFD session administratively down, for maintenance say, GoBGP tore down the whole BGP session. RFC 5882 says it must not. Now BGP stays up.
- #3798mergedChanging BFD timers restarted the BFD session and flapped BGP. Now the new timers are negotiated on the running session.
- #3785mergedDisabling BFD on a neighbor just stopped the packets, so the peer timed out and reset BGP over a link that was fine. Now GoBGP sends AdminDown first, as RFC 5880 says, and the peer keeps BGP up.
A multipath VPN over QUIC, and the QUIC library under it. Most of these add path MTU discovery end to end and take the tunnel MTU on a 1500-byte path from 1382 to 1418, with no fragmentation.
- mqvpn #430mergedPath MTU discovery (a PMTUD option), so the tunnel finds the largest packet the path really carries.Read the write-up
- mqvpn #449mergedWhen one client couldn't keep up, the server stopped reading the TUN that every client shares, so the others stalled too. Now only that client's packets are dropped. In a two-client test the other client went from about 18% ping loss to none.
Show 9 more
- xquic #110mergedIf the path shrank by just a few bytes, only full packets carrying an ACK were lost, and acks for the smaller ones kept hiding it, so black-hole detection never fired. Now an ack only clears a loss if its packet was at least as large.
- xquic #100mergedBlack-hole detection for PMTUD (RFC 8899 §4.3): if big probes start vanishing, it drops back down instead of stalling.Read the write-up
- xquic #101mergedPaths that are not validated yet no longer shrink the packet size for the whole connection.
- xquic #93A lost PMTUD probe no longer counts as congestion, so CUBIC stops cutting its speed every time a probe turns out too big. In my tests CUBIC with PMTUD on went from 27.5 back to 89.8 Mbit/s. Merged with my tests in it.Read the write-up
- mqvpn #428mergedKeeps the reorder header out of the TUN MTU when the MSS is updated.
- mqvpn #427mergedAccepts ICMP errors that come from our own address on Linux, which PMTUD needs.Read the write-up
- mqvpn #409mergedUses the new 1436-byte packet cap on a 1500-byte path, which raises the tunnel MTU from 1402 to 1418. Tested at about 3 Gbit/s each way with zero fragmentation.Read the write-up
- mqvpn #391mergedAdds a PathMTU setting, so the QUIC packet size and tunnel MTU follow the real path instead of a fixed size.Read the write-up
- xquic #86mergedxquic let an ACK frame spill past its packet size cap, so packets meant to fit a 1500-byte path came out at around 1458 bytes and got fragmented. The fix keeps a spilled ACK within the cap.Read the write-up
The reference compiler for P4, a language for programming packet processing. Its eBPF backend turns P4 into programs for the Linux kernel.
- #5914mergedRe-enables two eBPF tests skipped on newer kernels since 2025. One was a strict-aliasing miscompile of the generated C: clang 18 dropped stores to narrow key fields, so table lookups never matched.
- #5906mergedThe eBPF backend tests only ran on Ubuntu 22.04. Fixed each thing that broke on 24.04 with its newer clang and kernel, and moved CI to 24.04.
The first-boot provisioning tool for Fedora CoreOS, Flatcar and RHCOS. Butane turns readable YAML into its configs.
- #2348mergedA follow-up: files from local: and trees: had the same gap, so those are checked now too, and the error points at the exact file entry.
- #2343mergedButane accepted gzip compression on inline file contents without checking they were really gzip, so a config validated cleanly and then failed at first boot. It now checks the contents and reports the mistake when the config is built. I had reported it first, in coreos/butane#730.
Runtime-variable support for the auth_jwt_enabled directive, which makes JWT enforcement conditional on request context rather than fixed per location.
- #152mergedEnables integration with the nginx geo and map modules, for cases such as VPN subnet allowlisting and selective authentication bypass. It runs in production as the reverse proxy image on my own cluster.Read the write-up
Root-caused a GUI hang on the Disks page and fed the analysis into the upstream fix.
- #6224reportedTraced the hang through the Perl disk-management API to blocking SMART queries against USB-attached SSDs with no timeout handling, and supplied reproduction steps for the timeout patch.Read the write-up