Brad Fitzpatrick
8e1fbc4740
Previously the DERP handler served each connection for its lifetime on its hijacked connection's net/http handler goroutine. That goroutine's conn.serve stack frame kept the dead HTTP/1 server state reachable for the whole DERP connection: the http.conn and its 4KB bufio.Writer (hijack hands over c.bufw but conn.serve still references it, so derpserver returning it to its flush pool never made it collectable), the 4KB bufio.Reader, the upgrade *http.Request with its parsed headers, and the request context chain. The goroutine also kept the stack growth from the TLS handshake and HTTP request parsing. Instead, hand the connection off to a new goroutine and return from the handler (ala tailscale/corp@dc09e27aef), letting all the HTTP upgrade state be collected. Give Accept a smaller 1KB frame reader, draining and releasing the hijacked reader if it contains buffered bytes from a fast-start client, and a nil bufio.Writer so writes go through pooled buffers held only for the duration of a write instead of a per-connection buffer. Because the handler now returns at handoff time, cmd/derper's gauge_derper_tls_active_version decrement can no longer be deferred to handler return: intercept Hijack in the TLS metrics wrapper and, for hijacked connections (DERP, its WebSocket flavor, and CONNECT), decrement the gauge once when the hijacked connection closes, restoring the gauge's connection-lifetime semantics. Teach derpserver's TCP RTT stats to unwrap the close-hook conn so they still find the underlying *net.TCPConn. Also soften the UntypedHexString deprecation notices in types/key to warnings: the untyped hex string format is the DERP wire protocol's key encoding, so these call sites are legitimate and permanent, and a Deprecated marker just makes them light up in editors and linters. The cautionary text about the format's risks remains. Measured with 100,000 idle TLS DERP connections on linux/amd64: standing memory drops from 55.6KB to 32.6KB per connection (-41%). Updates #21064 Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com> Change-Id: If05a0c6ea79134807e9e8872861db216
DERP
This is the code for the Tailscale DERP server.
In general, you should not need to or want to run this code. The overwhelming majority of Tailscale users (both individuals and companies) do not.
In the happy path, Tailscale establishes direct connections between peers and data plane traffic flows directly between them, without using DERP for more than acting as a low bandwidth side channel to bootstrap the NAT traversal. If you find yourself wanting DERP for more bandwidth, the real problem is usually the network configuration of your Tailscale node(s), making sure that Tailscale can get direction connections via some mechanism.
If you've decided or been advised to run your own derper, then read on.
Caveats
-
Node sharing and other cross-Tailnet features don't work when using custom DERP servers.
-
DERP servers only see encrypted WireGuard packets and thus are not useful for network-level debugging.
-
The Tailscale control plane does certain geo-level steering features and optimizations that are not available when using custom DERP servers.
Guide to running cmd/derper
-
You must build and update the
cmd/derperbinary yourself. There are no packages. Usego install tailscale.com/cmd/derper@latestwith the latest version of Go. You should update this binary approximately as regularly as you update Tailscale nodes. If using--verify-clients, thederperbinary andtailscaledbinary on the machine must be built from the same git revision. (It might work otherwise, but they're developed and only tested together.) -
The DERP protocol does a protocol switch inside TLS from HTTP to a custom bidirectional binary protocol. It is thus incompatible with many HTTP proxies. Do not put
derperbehind another HTTP proxy. -
The
tailscaledclient does its own selection of the fastest/nearest DERP server based on latency measurements. Do not putderperbehind a global load balancer. -
DERP servers should ideally have both a static IPv4 and static IPv6 address. Both of those should be listed in the DERP map so the client doesn't need to rely on its DNS which might be broken and dependent on DERP to get back up.
-
A DERP server should not share an IP address with any other DERP server.
-
Avoid having multiple DERP nodes in a region. If you must, they all need to be meshed with each other and monitored. Having two one-node "regions" in the same datacenter is usually easier and more reliable than meshing, at the cost of more required connections from clients in some cases. If your clients aren't mobile (battery constrained), one node regions are definitely preferred. If you really need multiple nodes in a region for HA reasons, two is sufficient.
-
Monitor your DERP servers with
cmd/derpprobe. -
If using
--verify-clients, atailscaledmust be running alongside thederper, and all clients must be visible to the derper tailscaled in the ACL. -
If using
--verify-clients, atailscaledmust also be running alongside yourderpprobe, andderpprobeneeds to use--derp-map=local. -
The firewall on the
derpershould permit TCP ports 80 and 443 and UDP port 3478. -
Only LetsEncrypt certs are rotated automatically. Other cert updates require a restart.
-
Don't use a firewall in front of
derperthat suppressesRSTs upon receiving traffic to a dead or unknown connection. -
Don't rate-limit UDP STUN packets.
-
Don't rate-limit outbound TCP traffic (only inbound).
Diagnostics
This is not a complete guide on DERP diagnostics.
Running your own DERP services requires exeprtise in multi-layer network and application diagnostics. As the DERP runs multiple protocols at multiple layers and is not a regular HTTP(s) server you will need expertise in correlative analysis to diagnose the most tricky problems. There is no "plain text" or "open" mode of operation for DERP.
-
The debug handler is accessible at URL path
/debug/. It is only accessible over localhost or from a Tailscale IP address. -
Go pprof can be accessed via the debug handler at
/debug/pprof/ -
Prometheus compatible metrics can be gathered from the debug handler at
/debug/varz. -
cmd/stuncin the Tailscale repository provides a basic tool for diagnosing issues with STUN. -
cmd/derpprobeprovides a service for monitoring DERP cluster health. -
tailscale debug derpandtailscale netcheckprovide additional client driven diagnostic information for DERP communications. -
Tailscale logs may provide insight for certain problems, such as if DERPs are unreachable or peers are regularly not reachable in their DERP home regions. There are many possible misconfiguration causes for these problems, but regular log entries are a good first indicator that there is a problem.