tailscale/derp/derp.go

232 lines
7.6 KiB
Go
Raw Normal View History

// Copyright (c) 2020 Tailscale Inc & AUTHORS All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
// Package derp implements DERP, the Detour Encrypted Routing Protocol.
//
// DERP routes packets to clients using curve25519 keys as addresses.
//
// DERP is used by Tailscale nodes to proxy encrypted WireGuard
// packets through the Tailscale cloud servers when a direct path
// cannot be found or opened. DERP is a last resort. Both sides
// between very aggressive NATs, firewalls, no IPv6, etc? Well, DERP.
package derp
import (
"bufio"
"encoding/binary"
"errors"
"fmt"
"io"
"io/ioutil"
"time"
)
// MaxPacketSize is the maximum size of a packet sent over DERP.
// (This only includes the data bytes visible to magicsock, not
// including its on-wire framing overhead)
const MaxPacketSize = 64 << 10
// magic is the DERP magic number, sent in the frameServerKey frame
// upon initial connection.
const magic = "DERP🔑" // 8 bytes: 0x44 45 52 50 f0 9f 94 91
const (
nonceLen = 24
frameHeaderLen = 1 + 4 // frameType byte + 4 byte length
keyLen = 32
maxInfoLen = 1 << 20
keepAlive = 60 * time.Second
)
// ProtocolVersion is bumped whenever there's a wire-incompatible change.
// * version 1 (zero on wire): consistent box headers, in use by employee dev nodes a bit
// * version 2: received packets have src addrs in frameRecvPacket at beginning
const ProtocolVersion = 2
// frameType is the one byte frame type at the beginning of the frame
// header. The second field is a big-endian uint32 describing the
// length of the remaining frame (not including the initial 5 bytes).
type frameType byte
/*
Protocol flow:
Login:
* client connects
* server sends frameServerKey
* client sends frameClientInfo
* server sends frameServerInfo
Steady state:
* server occasionally sends frameKeepAlive (or framePing)
* client responds to any framePing with a framePong
* client sends frameSendPacket
* server then sends frameRecvPacket to recipient
*/
const (
frameServerKey = frameType(0x01) // 8B magic + 32B public key + (0+ bytes future use)
frameClientInfo = frameType(0x02) // 32B pub key + 24B nonce + naclbox(json)
frameServerInfo = frameType(0x03) // 24B nonce + naclbox(json)
frameSendPacket = frameType(0x04) // 32B dest pub key + packet bytes
frameForwardPacket = frameType(0x0a) // 32B src pub key + 32B dst pub key + packet bytes
frameRecvPacket = frameType(0x05) // v0/1: packet bytes, v2: 32B src pub key + packet bytes
frameKeepAlive = frameType(0x06) // no payload, no-op (to be replaced with ping/pong)
frameNotePreferred = frameType(0x07) // 1 byte payload: 0x01 or 0x00 for whether this is client's home node
// framePeerGone is sent from server to client to signal that
// a previous sender is no longer connected. That is, if A
// sent to B, and then if A disconnects, the server sends
// framePeerGone to B so B can forget that a reverse path
// exists on that connection to get back to A.
framePeerGone = frameType(0x08) // 32B pub key of peer that's gone
// framePeerPresent is like framePeerGone, but for other
// members of the DERP region when they're meshed up together.
framePeerPresent = frameType(0x09) // 32B pub key of peer that's connected
// frameWatchConns is how one DERP node in a regional mesh
// subscribes to the others in the region.
// There's no payload. If the sender doesn't have permission, the connection
// is closed. Otherwise, the client is initially flooded with
// framePeerPresent for all connected nodes, and then a stream of
// framePeerPresent & framePeerGone has peers connect and disconnect.
frameWatchConns = frameType(0x10)
// frameClosePeer is a privileged frame type (requires the
// mesh key for now) that closes the provided peer's
// connection. (To be used for cluster load balancing
// purposes, when clients end up on a non-ideal node)
frameClosePeer = frameType(0x11) // 32B pub key of peer to close.
framePing = frameType(0x12) // 8 byte ping payload, to be echoed back in framePong
framePong = frameType(0x13) // 8 byte payload, the contents of the ping being replied to
// frameHealth is sent from server to client to tell the client
// if their connection is unhealthy somehow. Currently the only unhealthy state
// is whether the connection is detected as a duplicate.
// The entire frame body is the text of the error message. An empty message
// clears the error state.
frameHealth = frameType(0x14)
// frameRestarting is sent from server to client for the
// server to declare that it's restarting. Payload is two big
// endian uint32 durations in milliseconds: when to reconnect,
// and how long to try total. See ServerRestartingMessage docs for
// more details on how the client should interpret them.
frameRestarting = frameType(0x15)
)
var bin = binary.BigEndian
func writeUint32(bw *bufio.Writer, v uint32) error {
var b [4]byte
bin.PutUint32(b[:], v)
derp: make writeUint32 and readUint32 not allocate The allocations are small, but they're easy enough to avoid. And it removes some clutter from the pprof output. name old time/op new time/op delta SendRecv/msgsize=10-8 10.1µs ± 9% 9.7µs ± 7% -3.45% (p=0.035 n=14+14) SendRecv/msgsize=100-8 8.12µs ± 7% 7.38µs ± 9% -9.02% (p=0.000 n=15+15) SendRecv/msgsize=1000-8 9.51µs ±25% 8.76µs ±22% ~ (p=0.202 n=15+15) SendRecv/msgsize=10000-8 21.1µs ±25% 19.9µs ±14% ~ (p=0.270 n=15+14) WriteUint32-8 25.1ns ± 4% 21.3ns ±12% -15.01% (p=0.000 n=14+14) ReadUint32-8 35.4ns ± 4% 21.9ns ± 4% -38.06% (p=0.000 n=15+15) name old alloc/op new alloc/op delta SendRecv/msgsize=10-8 182B ± 2% 169B ± 1% -7.22% (p=0.000 n=15+13) SendRecv/msgsize=100-8 282B ± 1% 265B ± 1% -5.85% (p=0.000 n=15+15) SendRecv/msgsize=1000-8 1.19kB ± 1% 1.18kB ± 0% -1.26% (p=0.000 n=14+15) SendRecv/msgsize=10000-8 19.3kB ± 4% 18.7kB ± 4% -3.44% (p=0.006 n=12+12) WriteUint32-8 4.00B ± 0% 0.00B -100.00% (p=0.000 n=15+15) ReadUint32-8 4.00B ± 0% 0.00B -100.00% (p=0.000 n=15+15) name old allocs/op new allocs/op delta SendRecv/msgsize=10-8 8.00 ± 0% 4.00 ± 0% -50.00% (p=0.000 n=15+15) SendRecv/msgsize=100-8 8.00 ± 0% 4.00 ± 0% -50.00% (p=0.000 n=15+15) SendRecv/msgsize=1000-8 8.00 ± 0% 4.00 ± 0% -50.00% (p=0.000 n=15+15) SendRecv/msgsize=10000-8 8.47 ±17% 5.00 ± 0% -40.94% (p=0.000 n=15+14) WriteUint32-8 1.00 ± 0% 0.00 -100.00% (p=0.000 n=15+15) ReadUint32-8 1.00 ± 0% 0.00 -100.00% (p=0.000 n=15+15) Signed-off-by: Josh Bleecher Snyder <josharian@gmail.com>
2020-08-08 00:25:59 +01:00
// Writing a byte at a time is a bit silly,
// but it causes b not to escape,
// which more than pays for the silliness.
for _, c := range &b {
err := bw.WriteByte(c)
if err != nil {
return err
}
}
return nil
}
func readUint32(br *bufio.Reader) (uint32, error) {
derp: make writeUint32 and readUint32 not allocate The allocations are small, but they're easy enough to avoid. And it removes some clutter from the pprof output. name old time/op new time/op delta SendRecv/msgsize=10-8 10.1µs ± 9% 9.7µs ± 7% -3.45% (p=0.035 n=14+14) SendRecv/msgsize=100-8 8.12µs ± 7% 7.38µs ± 9% -9.02% (p=0.000 n=15+15) SendRecv/msgsize=1000-8 9.51µs ±25% 8.76µs ±22% ~ (p=0.202 n=15+15) SendRecv/msgsize=10000-8 21.1µs ±25% 19.9µs ±14% ~ (p=0.270 n=15+14) WriteUint32-8 25.1ns ± 4% 21.3ns ±12% -15.01% (p=0.000 n=14+14) ReadUint32-8 35.4ns ± 4% 21.9ns ± 4% -38.06% (p=0.000 n=15+15) name old alloc/op new alloc/op delta SendRecv/msgsize=10-8 182B ± 2% 169B ± 1% -7.22% (p=0.000 n=15+13) SendRecv/msgsize=100-8 282B ± 1% 265B ± 1% -5.85% (p=0.000 n=15+15) SendRecv/msgsize=1000-8 1.19kB ± 1% 1.18kB ± 0% -1.26% (p=0.000 n=14+15) SendRecv/msgsize=10000-8 19.3kB ± 4% 18.7kB ± 4% -3.44% (p=0.006 n=12+12) WriteUint32-8 4.00B ± 0% 0.00B -100.00% (p=0.000 n=15+15) ReadUint32-8 4.00B ± 0% 0.00B -100.00% (p=0.000 n=15+15) name old allocs/op new allocs/op delta SendRecv/msgsize=10-8 8.00 ± 0% 4.00 ± 0% -50.00% (p=0.000 n=15+15) SendRecv/msgsize=100-8 8.00 ± 0% 4.00 ± 0% -50.00% (p=0.000 n=15+15) SendRecv/msgsize=1000-8 8.00 ± 0% 4.00 ± 0% -50.00% (p=0.000 n=15+15) SendRecv/msgsize=10000-8 8.47 ±17% 5.00 ± 0% -40.94% (p=0.000 n=15+14) WriteUint32-8 1.00 ± 0% 0.00 -100.00% (p=0.000 n=15+15) ReadUint32-8 1.00 ± 0% 0.00 -100.00% (p=0.000 n=15+15) Signed-off-by: Josh Bleecher Snyder <josharian@gmail.com>
2020-08-08 00:25:59 +01:00
var b [4]byte
// Reading a byte at a time is a bit silly,
// but it causes b not to escape,
// which more than pays for the silliness.
for i := range &b {
c, err := br.ReadByte()
if err != nil {
return 0, err
}
b[i] = c
}
derp: make writeUint32 and readUint32 not allocate The allocations are small, but they're easy enough to avoid. And it removes some clutter from the pprof output. name old time/op new time/op delta SendRecv/msgsize=10-8 10.1µs ± 9% 9.7µs ± 7% -3.45% (p=0.035 n=14+14) SendRecv/msgsize=100-8 8.12µs ± 7% 7.38µs ± 9% -9.02% (p=0.000 n=15+15) SendRecv/msgsize=1000-8 9.51µs ±25% 8.76µs ±22% ~ (p=0.202 n=15+15) SendRecv/msgsize=10000-8 21.1µs ±25% 19.9µs ±14% ~ (p=0.270 n=15+14) WriteUint32-8 25.1ns ± 4% 21.3ns ±12% -15.01% (p=0.000 n=14+14) ReadUint32-8 35.4ns ± 4% 21.9ns ± 4% -38.06% (p=0.000 n=15+15) name old alloc/op new alloc/op delta SendRecv/msgsize=10-8 182B ± 2% 169B ± 1% -7.22% (p=0.000 n=15+13) SendRecv/msgsize=100-8 282B ± 1% 265B ± 1% -5.85% (p=0.000 n=15+15) SendRecv/msgsize=1000-8 1.19kB ± 1% 1.18kB ± 0% -1.26% (p=0.000 n=14+15) SendRecv/msgsize=10000-8 19.3kB ± 4% 18.7kB ± 4% -3.44% (p=0.006 n=12+12) WriteUint32-8 4.00B ± 0% 0.00B -100.00% (p=0.000 n=15+15) ReadUint32-8 4.00B ± 0% 0.00B -100.00% (p=0.000 n=15+15) name old allocs/op new allocs/op delta SendRecv/msgsize=10-8 8.00 ± 0% 4.00 ± 0% -50.00% (p=0.000 n=15+15) SendRecv/msgsize=100-8 8.00 ± 0% 4.00 ± 0% -50.00% (p=0.000 n=15+15) SendRecv/msgsize=1000-8 8.00 ± 0% 4.00 ± 0% -50.00% (p=0.000 n=15+15) SendRecv/msgsize=10000-8 8.47 ±17% 5.00 ± 0% -40.94% (p=0.000 n=15+14) WriteUint32-8 1.00 ± 0% 0.00 -100.00% (p=0.000 n=15+15) ReadUint32-8 1.00 ± 0% 0.00 -100.00% (p=0.000 n=15+15) Signed-off-by: Josh Bleecher Snyder <josharian@gmail.com>
2020-08-08 00:25:59 +01:00
return bin.Uint32(b[:]), nil
}
func readFrameTypeHeader(br *bufio.Reader, wantType frameType) (frameLen uint32, err error) {
gotType, frameLen, err := readFrameHeader(br)
if err == nil && wantType != gotType {
err = fmt.Errorf("bad frame type 0x%X, want 0x%X", gotType, wantType)
}
return frameLen, err
}
func readFrameHeader(br *bufio.Reader) (t frameType, frameLen uint32, err error) {
tb, err := br.ReadByte()
if err != nil {
return 0, 0, err
}
frameLen, err = readUint32(br)
if err != nil {
return 0, 0, err
}
return frameType(tb), frameLen, nil
}
// readFrame reads a frame header and then reads its payload into
// b[:frameLen].
//
// If the frame header length is greater than maxSize, readFrame returns
// an error after reading the frame header.
//
// If the frame is less than maxSize but greater than len(b), len(b)
// bytes are read, err will be io.ErrShortBuffer, and frameLen and t
// will both be set. That is, callers need to explicitly handle when
// they get more data than expected.
func readFrame(br *bufio.Reader, maxSize uint32, b []byte) (t frameType, frameLen uint32, err error) {
t, frameLen, err = readFrameHeader(br)
if err != nil {
return 0, 0, err
}
if frameLen > maxSize {
return 0, 0, fmt.Errorf("frame header size %d exceeds reader limit of %d", frameLen, maxSize)
}
n, err := io.ReadFull(br, b[:minUint32(frameLen, uint32(len(b)))])
if err != nil {
return 0, 0, err
}
remain := frameLen - uint32(n)
if remain > 0 {
if _, err := io.CopyN(ioutil.Discard, br, int64(remain)); err != nil {
return 0, 0, err
}
err = io.ErrShortBuffer
}
return t, frameLen, err
}
func writeFrameHeader(bw *bufio.Writer, t frameType, frameLen uint32) error {
if err := bw.WriteByte(byte(t)); err != nil {
return err
}
return writeUint32(bw, frameLen)
}
// writeFrame writes a complete frame & flushes it.
func writeFrame(bw *bufio.Writer, t frameType, b []byte) error {
if len(b) > 10<<20 {
return errors.New("unreasonably large frame write")
}
if err := writeFrameHeader(bw, t, uint32(len(b))); err != nil {
return err
}
if _, err := bw.Write(b); err != nil {
return err
}
return bw.Flush()
}
func minUint32(a, b uint32) uint32 {
if a < b {
return a
}
return b
}