Using an exit node could destroy direct peer connections: the agent's UID
bypass protected outbound routing, but reverse-path checks look up replies
without that UID and selected the overlay default. A real-host A/B check
confirmed that relaxing only strict reverse-path filtering restored direct
connections and exit HTTP traffic; the diagnostic host changes were removed.
Keep strict filtering and protect the Linux underlay with a dedicated masked
mark bit, 0x40000000. Mark the dedicated service UID in OUTPUT, save only that
bit in conntrack and restore it before PREROUTING reverse-path checks. Install
IPv4/IPv6 mark and UID main-table bypass rules before Tailscale's priorities;
leave application-specific and exit-default routing after them. Preserve
foreign mark bits, main-table routes and foreign firewall/policy rules.
Protect the first inbound TCP/UDP packet using bounded, locally owned socket
addresses/ports from /proc and an actual socket match. Refresh socket protection
before sending private transport negotiations and on discovery ticks, without
firewall writes for an unchanged socket set. Remote announcements never become
OS arguments. A failed refresh leaves active exit routing in place. Install
protection before routes, remove routes before protection, roll back partial
activation, and remove only exact interface tags, including legacy UID rules.
Reject strict kernel rp_filter without src_valid_mark with an actionable error;
never weaken filtering or change host sysctls. Document mark-aware Netfilter
requirements and provide the trusted ip6tables path in the NixOS module.
Other platforms retain their existing exit implementation. Peer wire formats,
identity derivation, local IPC and persistent storage are unchanged.
Make the tray devices grid fill and shrink with the window, constrain the first
cell and give remaining width to traffic plots. Keep displayed end-to-end
protocol and its short encryption tooltip consistent, independent of carrier
protocol. Move handshake/session age into its own column beside the existing
iroh control RTT. Use stable per-device hover identities and concise tooltips;
missing protocols/measurements remain explicit rather than invented.
Validation on Linux through nix develop:
- cargo fmt --all -- --check
- cargo clippy --locked --workspace --all-targets -- -D warnings
- cargo test --locked --workspace --all-targets: 529 passed, 3 ignored;
inherited TSUNAGI_CONTROL_SOCKET and TSUNAGI_CONTROL_GROUP were unset
- explicit privileged kernel regression: 1 passed, using disposable namespaces,
real TCP/UDP IPv4/IPv6, strict kernel and Netfilter RPF, first inbound packets,
dynamic UDP binding, application exit routing/IPv6 blocking, repeated apply
and exact cleanup with foreign Tailscale-priority rules retained
- headless egui grid/scroll-area rendering at multiple sizes and control RTT,
mark masks/order, passive socket parsing, partial failure and cleanup tests
- git diff --check
Earlier iterations caught rtnetlink family API and nonexhaustive initializer
errors, Clippy warnings, egui's implicit minimum column width, and kernel-test
fixture issues (wildcard echo source, blocked IPv6 neighbour discovery and a
closed child stdout pipe). Corrected before final checks; the stdout-only test
fix was checked with fmt, Clippy and the explicit kernel regression afterward.
Native tray rendering and macOS/Windows checks were not run. The running host
agent was not replaced or restarted. Requested release tag: v0.1.0-rc.32.
Choosing an exit node on Windows did nothing: the using side was refused,
because Windows has no per-user routing and the agent's own traffic would loop
into the tunnel it carries.
Now the overlay interface gets 0.0.0.0/1 and 128.0.0.0/1 routes (and ::/1,
8000::/1 as an IPv6 blackhole), and every underlay address the agent talks to
directly gets a /32 route through the physical default gateway, marked with
route metric 7373 so stale ones are found and removed. Bypass routes are
written before the halves and removed after them; a failed step removes what
was written. All routes are in the active store.
The addresses come from the runtime: each network runtime hands the interface
the verified direct IPv4 paths of its control connections and the relay hosts
in use (Interface::set_underlay, refreshed every dial round); relay names are
resolved when the rules are applied. ExitHostRules::needs_bypass (true only on
Windows) gates it, so Linux and macOS plans never carry or re-apply on path
changes. ExitHostPlan gains `bypass` and `bypass_hosts`; networks leaving
forget their entry.
Limits: a path that appears after the last apply goes through the tunnel until
the next apply; data-plane-only addresses are not collected separately (the
data links ride the same iroh paths). Nothing was run on Windows.
Checks: cargo fmt --check, cargo clippy --locked --workspace --all-targets
-D warnings and cargo test --locked --workspace --all-targets passed on macOS;
scripts and ordering are unit-tested with an injected runner and resolver, and
Interface::set_underlay with the mock host.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Windows reported exit nodes as unsupported. The offering side now creates one
WinNAT object per offered range (New-NetNat, named tsunagi-exit:<if>:<range>),
prunes this interface's stale ones, reads (never writes) IPv4 forwarding of
the overlay interface, and removes everything on clear. All values are
single-quoted PowerShell literals.
The using side is deliberately not implemented: Windows has no per-user or
per-process routing, so the agent's own QUIC/relay traffic cannot be kept out
of a default route through the overlay and would loop. It reports that reason
and installs no route. The CLI/tray forwarding hint and the kernel-forwarding
warning name the Windows command; docs/exit-node.md has a Windows section.
Checks: cargo fmt --check, cargo clippy --locked --workspace --all-targets
-D warnings and cargo test --locked --workspace --all-targets passed on macOS.
The script builders and plan logic are unit-tested with an injected runner on
every platform; nothing was run against a real Windows host or cross-compiled
(no Windows target installed).
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The exit-node host rules existed only for Linux (iptables + policy routing);
macOS reported "only supported on Linux yet".
Offering: a pf anchor com.apple/tsunagi-exit-<if> (evaluated by the stock
pf.conf) holds a nat rule per offered range and a pass-in rule. pf is enabled
if needed and the pfctl -E reference is released when the rules are removed.
net.inet.ip.forwarding is only read, never written, as on Linux.
Using: macOS has no uidrange policy routing and the agent is root, so the
overlay gets 0.0.0.0/1 and 128.0.0.0/1 routes and a pf route-to rule (matched
by the agent's uid) sends the agent's own QUIC/relay traffic back out the
physical interface, translated to its address. The rules are loaded before the
routes, so a failed load never sends traffic into a tunnel it cannot exempt
the agent from. IPv6 is blocked with rejecting ::/1 and 8000::/1 routes.
ExitHostPlan gains `overlay` (every network range) for the route-to exemption;
Linux ignores it.
Known limits: the egress is the default route's interface at apply time; other
root processes also bypass the tunnel. Forwarding hints in the CLI and tray
name the macOS sysctl. The rule text and command order are unit-tested with a
scripted host; nothing was run against a live pf/route (needs root), so the
route-to + nat second-pass behaviour is unverified.
Checks: cargo fmt --check, cargo clippy --locked --workspace --all-targets
-D warnings and cargo test --locked --workspace --all-targets, run on macOS
only.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The command line package is now `tsunagi` and builds the `tsng` executable,
which the packages, the unit, the installers and the workflows install and
run. The core library package becomes `tsunagi-core`, keeping the library name
`tsunagi`, so no code changes `use tsunagi::...`. Directories follow
(crates/tsunagi-core, crates/tsunagi). Every command shown to a person is now
`tsng ...`.
Co-Authored-By: Claude <noreply@anthropic.com>
IPv6 has no way through an exit node, so the same policy rules get an
unreachable default for IPv6 instead of leaking it. The tray shows a green
icon in the tile corner while an exit node is in use, with a one-line
confirmation to stop it; confirmations are shortened.
Co-Authored-By: Claude <noreply@anthropic.com>
A member can offer itself as an exit node per network (off by default); it is
announced over the control channel (ctrl/4) once its iptables masquerade and
forward rules are in place, with a warning when kernel forwarding is off. A
device picks one member as its exit node via policy routing, with the agent's
own uid exempt from the tunnel; if the exit node disappears the rules stay
and traffic is dropped, loudly reported. CLI: `network exit-node`,
`exit-node`; tray: tile checkbox, warnings and a per-device icon with
confirmation. Linux only.
Co-Authored-By: Claude <noreply@anthropic.com>
Aleksandr Bogomiakov
UltradesuandClaude Sonnet 5.5