Files
Brad Fitzpatrick 04ae61fe4b tstest/integration/jswasmtest: add headless-Chromium tests for @tailscale/connect
Add Go tests that drive a real headless Chromium (via chromedp) against
the built cmd/tsconnect/pkg/ artifact and verify the @tailscale/connect
public API surface end-to-end. The package has not been republished in
three years, in part because no test exercises the produced artifact at
runtime — only tsc --noEmit and a Go build run in CI.

TestCreateIPN loads pkg.js into the browser, calls createIPN with a junk
auth key, and asserts that pkg.createIPN / pkg.runSSHSession are
functions and that createIPN() returns an IPN with the documented
run/login/logout/ssh/fetch methods. No control-plane traffic.

TestFetchTailnetPeer stands up a full local tailnet (testcontrol +
DERP + a tsnet.Server peer) and verifies that the browser-side WASM
client can join over WebSocket-noise to the same control, connect to
DERP over WSS, and then ipn.fetch() an HTTP service hosted on the tsnet
peer through the tailnet. The test asserts the response body matches a
known string. Browser state transitions are logged: NoState -> NeedsLogin
-> Starting -> Running.

Tests are opt-in via --run-headless-browser-tests (matching the existing
--run-vm-tests pattern in tstest/natlab/vmtest) so they never fire in
casual `go test ./...` runs. When the flag is set, a test is skipped if
cmd/tsconnect/pkg/ has not been built, and fails with t.Error if no
chromium binary is found on $PATH (honoring $CHROME_BIN as an override).
findChromium also falls back to /Applications/Google Chrome.app and
/Applications/Chromium.app on darwin, since macOS Chrome's executable
lives inside an .app bundle and is not on $PATH by default. The
.github/workflows/test.yml wasm job is extended to install
google-chrome-stable and run the tests with the flag after build-pkg.

To prevent silently testing a stale pkg/main.wasm (built from an older
checkout than the rest of the test invocation), build-pkg now writes
pkg/build-info.json recording the sha256 of the raw (pre-wasm-opt)
go-build output. The test does its own `go build` of
cmd/tsconnect/wasm with the same -tags/-trimpath/-ldflags (factored
into a new cmd/tsconnect/wasmbuild package shared by both call sites)
and t.Fatalfs with a "rebuild" instruction on mismatch. Cost is
near-zero because the Go build cache from the prior build-pkg makes
the rebuild a cache hit.

The new wasmbuild package also replaces cmd/tsconnect's hardcoded -tags
string with a minimal-feature-set computation. wasmbuild.Keep names the
small set of feature/featuretags entries the browser client actually
needs (netstack, logtail, dns, health, c2n, ipnbus); wasmbuild.Tags()
emits a ts_omit_<f> for every other
omittable feature in feature/featuretags.Features, with transitive deps
expanded via featuretags.Requires. An init() panics if Keep references
a feature unknown to feature/featuretags so a rename there fails
loudly. Net effect on size: 32M raw / 9.4M brotli before this change,
25M raw / 4.4M brotli after — vs the last-published 1.39.98 at 21M /
3.8M. The transitive package-import graph is unchanged (176
tailscale.com/* packages either way): featuretags omits eliminate
dead code via `const HasX = false`, not imports. Trimming the import
graph would require a separate, larger refactor splitting interface
packages by build tag.

Writing TestFetchTailnetPeer surfaced several real issues, all fixed
here:

  * cmd/tsconnect built the wasm with the nethttpomithttp2 tag, but
    control/ts2021 (since commit 1d93bdce2, "control/controlclient:
    remove x/net/http2, use net/http", Oct 2025) requires HTTP/2 from
    net/http's bundled implementation. With nethttpomithttp2 set, the
    bundle is excluded and the wasm client cannot speak HTTP/2 to any
    control plane, including production. Drop the tag. Wasm size grows
    ~1 MB raw / ~300 KB brotli (more than offset by the feature
    pruning above). The last published @tailscale/connect (1.39.98,
    early 2023) pre-dates the regression, which is why no consumer has
    reported the breakage.

  * tstest/integration/testcontrol.Server's /ts2021 noise upgrade
    endpoint rejected anything but POST. WebSocket clients (the only
    transport available to browser-WASM) come in as GET. Allow both;
    the controlhttp AcceptHTTP path dispatches on the Upgrade header,
    so the websocket library still enforces GET for WS upgrades.
    This matches production, where the same controlhttpserver.AcceptHTTP
    routes purely on the Upgrade header without checking method.

  * derp/derphttp's urlString built the DERP URL from node.HostName
    only, dropping node.DERPPort. Non-WS clients use a separate code
    path (connectToHost) that honors DERPPort, but WebSocket-only
    clients (browser-WASM) went through urlString and so could not
    reach a DERP running on any port other than 443. Include the port
    when it differs from the scheme default.

Also move addWebSocketSupport from cmd/derper (where it was main-only)
to derp/derpserver.AddWebSocketSupport so tstest/integration.RunDERPAndSTUN
can wrap its DERP handler with WebSocket support — without that, the
test DERP would not accept the browser's wss connection.

Fixes #9394

Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Change-Id: Iff9cdee303e3b239924249b5bffb2fd04e02f391
2026-05-20 10:48:29 -07:00
..
2026-04-07 16:14:26 -07:00
2023-05-02 13:42:25 -07:00

DERP

This directory (and subdirectories) contain the DERP code. The server itself is in ../cmd/derper.

DERP is a packet relay system (client and servers) where peers are addressed using WireGuard public keys instead of IP addresses.

It relays two types of packets:

  • "Disco" discovery messages (see ../disco) as the a side channel during NAT traversal.

  • Encrypted WireGuard packets as the fallback of last resort when UDP is blocked or NAT traversal fails.

DERP Map

Each client receives a "DERP Map" from the coordination server describing the DERP servers the client should try to use.

The client picks its home "DERP home" based on latency. This is done to keep costs low by avoid using cloud load balancers (pricey) or anycast, which would necessarily require server-side routing between DERP regions.

Clients pick their DERP home and report it to the coordination server which shares it to all the peers in the tailnet. When a peer wants to send a packet and it doesn't already have a WireGuard session open, it sends disco messages (some direct, and some over DERP), trying to do the NAT traversal. The client will make connections to multiple DERP regions as needed. Only the DERP home region connection needs to be alive forever.

DERP Regions

Tailscale runs 1 or more DERP nodes (instances of cmd/derper) in various geographic regions to make sure users have low latency to their DERP home.

Regions generally have multiple nodes per region "meshed" (routing to each other) together for redundancy: it allows for cloud failures or upgrades without kicking users out to a higher latency region. Instead, clients will reconnect to the next node in the region. Each node in the region is required to to be meshed with every other node in the region and forward packets to the other nodes in the region. Packets are forwarded only one hop within the region. There is no routing between regions. The assumption is that the mesh TCP connections are over a VPC that's very fast, low latency, and not charged per byte. The coordination server assigns the list of nodes in a region as a function of the tailnet, so all nodes within a tailnet should generally be on the same node and not require forwarding. Only after a failure do clients of a particular tailnet get split between nodes in a region and require inter-node forwarding. But over time it balances back out. There's also an admin-only DERP frame type to force close the TCP connection of a particular client to force them to reconnect to their primary if the operator wants to force things to balance out sooner. (Using the (*derphttp.Client).ClosePeer method, as used by Tailscale's internal rarely-used cmd/derpprune maintenance tool)

We generally run a minimum of three nodes in a region not for quorum reasons (there's no voting) but just because two is too uncomfortably few for cascading failure reasons: if you're running two nodes at 51% load (CPU, memory, etc) and then one fails, that makes the second one fail. With three or more nodes, you can run each node a bit hotter.