tailscale/control/controlbase/conn_test.go
David Anderson 02ad987e24 control/controlbase: make the protocol version number selectable.
This is so that we can plumb our client capability version through
the protocol as the Noise version. The capability version increments
more frequently than strictly required (the Noise version only needs
to change when cryptographically-significant changes are made to
the protocol, whereas the capability version also indicates changes
in non-cryptographically-significant parts of the protocol), but this
gives us a safe pre-auth way to determine if the client supports
future protocol features, while still relying on Noise's strong
assurance that the client and server have agreed on the same version.

Currently, the server executes the same protocol regardless of the
version number, and just presents the version to the caller so they
can do capability-based things in the upper RPC protocol. In future,
we may add a ratchet to disallow obsolete protocols, or vary the
Noise handshake behavior based on requested version.

Updates #3488

Signed-off-by: David Anderson <danderson@tailscale.com>
2022-04-07 13:25:28 -07:00

419 lines
10 KiB
Go

// Copyright (c) 2021 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 controlbase
import (
"bufio"
"bytes"
"context"
"crypto/rand"
"encoding/binary"
"fmt"
"io"
"net"
"runtime"
"strings"
"sync"
"testing"
"testing/iotest"
"time"
chp "golang.org/x/crypto/chacha20poly1305"
"golang.org/x/net/nettest"
tsnettest "tailscale.com/net/nettest"
"tailscale.com/types/key"
)
const testProtocolVersion = 1
func TestMessageSize(t *testing.T) {
// This test is a regression guard against someone looking at
// maxCiphertextSize, going "huh, we could be more efficient if it
// were larger, and accidentally violating the Noise spec. Do not
// change this max value, it's a deliberate limitation of the
// cryptographic protocol we use (see Section 3 "Message Format"
// of the Noise spec).
const max = 65535
if maxCiphertextSize > max {
t.Fatalf("max ciphertext size is %d, which is larger than the maximum noise message size %d", maxCiphertextSize, max)
}
}
func TestConnBasic(t *testing.T) {
client, server := pair(t)
sb := sinkReads(server)
want := "test"
if _, err := io.WriteString(client, want); err != nil {
t.Fatalf("client write failed: %v", err)
}
client.Close()
if got := sb.String(4); got != want {
t.Fatalf("wrong content received: got %q, want %q", got, want)
}
if err := sb.Error(); err != io.EOF {
t.Fatal("client close wasn't seen by server")
}
if sb.Total() != 4 {
t.Fatalf("wrong amount of bytes received: got %d, want 4", sb.Total())
}
}
// bufferedWriteConn wraps a net.Conn and gives control over how
// Writes get batched out.
type bufferedWriteConn struct {
net.Conn
w *bufio.Writer
manualFlush bool
}
func (c *bufferedWriteConn) Write(bs []byte) (int, error) {
n, err := c.w.Write(bs)
if err == nil && !c.manualFlush {
err = c.w.Flush()
}
return n, err
}
// TestFastPath exercises the Read codepath that can receive multiple
// Noise frames at once and decode each in turn without making another
// syscall.
func TestFastPath(t *testing.T) {
s1, s2 := tsnettest.NewConn("noise", 128000)
b := &bufferedWriteConn{s1, bufio.NewWriterSize(s1, 10000), false}
client, server := pairWithConns(t, b, s2)
b.manualFlush = true
sb := sinkReads(server)
const packets = 10
s := "test"
for i := 0; i < packets; i++ {
// Many separate writes, to force separate Noise frames that
// all get buffered up and then all sent as a single slice to
// the server.
if _, err := io.WriteString(client, s); err != nil {
t.Fatalf("client write1 failed: %v", err)
}
}
if err := b.w.Flush(); err != nil {
t.Fatalf("client flush failed: %v", err)
}
client.Close()
want := strings.Repeat(s, packets)
if got := sb.String(len(want)); got != want {
t.Fatalf("wrong content received: got %q, want %q", got, want)
}
if err := sb.Error(); err != io.EOF {
t.Fatalf("client close wasn't seen by server")
}
}
// Writes things larger than a single Noise frame, to check the
// chunking on the encoder and decoder.
func TestBigData(t *testing.T) {
client, server := pair(t)
serverReads := sinkReads(server)
clientReads := sinkReads(client)
const sz = 15 * 1024 // 15KiB
clientStr := strings.Repeat("abcde", sz/5)
serverStr := strings.Repeat("fghij", sz/5*2)
if _, err := io.WriteString(client, clientStr); err != nil {
t.Fatalf("writing client>server: %v", err)
}
if _, err := io.WriteString(server, serverStr); err != nil {
t.Fatalf("writing server>client: %v", err)
}
if serverGot := serverReads.String(sz); serverGot != clientStr {
t.Error("server didn't receive what client sent")
}
if clientGot := clientReads.String(2 * sz); clientGot != serverStr {
t.Error("client didn't receive what server sent")
}
getNonce := func(n [chp.NonceSize]byte) uint64 {
if binary.BigEndian.Uint32(n[:4]) != 0 {
panic("unexpected nonce")
}
return binary.BigEndian.Uint64(n[4:])
}
// Reach into the Conns and verify the cipher nonces advanced as
// expected.
if getNonce(client.tx.nonce) != getNonce(server.rx.nonce) {
t.Error("desynchronized client tx nonce")
}
if getNonce(server.tx.nonce) != getNonce(client.rx.nonce) {
t.Error("desynchronized server tx nonce")
}
if n := getNonce(client.tx.nonce); n != 4 {
t.Errorf("wrong client tx nonce, got %d want 4", n)
}
if n := getNonce(server.tx.nonce); n != 8 {
t.Errorf("wrong client tx nonce, got %d want 8", n)
}
}
// readerConn wraps a net.Conn and routes its Reads through a separate
// io.Reader.
type readerConn struct {
net.Conn
r io.Reader
}
func (c readerConn) Read(bs []byte) (int, error) { return c.r.Read(bs) }
// Check that the receiver can handle not being able to read an entire
// frame in a single syscall.
func TestDataTrickle(t *testing.T) {
s1, s2 := tsnettest.NewConn("noise", 128000)
client, server := pairWithConns(t, s1, readerConn{s2, iotest.OneByteReader(s2)})
serverReads := sinkReads(server)
const sz = 10000
clientStr := strings.Repeat("abcde", sz/5)
if _, err := io.WriteString(client, clientStr); err != nil {
t.Fatalf("writing client>server: %v", err)
}
serverGot := serverReads.String(sz)
if serverGot != clientStr {
t.Error("server didn't receive what client sent")
}
}
func TestConnStd(t *testing.T) {
// You can run this test manually, and noise.Conn should pass all
// of them except for TestConn/PastTimeout,
// TestConn/FutureTimeout, TestConn/ConcurrentMethods, because
// those tests assume that write errors are recoverable, and
// they're not on our Conn due to cipher security.
t.Skip("not all tests can pass on this Conn, see https://github.com/golang/go/issues/46977")
nettest.TestConn(t, func() (c1 net.Conn, c2 net.Conn, stop func(), err error) {
s1, s2 := tsnettest.NewConn("noise", 4096)
controlKey := key.NewMachine()
machineKey := key.NewMachine()
serverErr := make(chan error, 1)
go func() {
var err error
c2, err = Server(context.Background(), s2, controlKey, testProtocolVersion, nil)
serverErr <- err
}()
c1, err = Client(context.Background(), s1, machineKey, controlKey.Public(), testProtocolVersion)
if err != nil {
s1.Close()
s2.Close()
return nil, nil, nil, fmt.Errorf("connecting client: %w", err)
}
if err := <-serverErr; err != nil {
c1.Close()
s1.Close()
s2.Close()
return nil, nil, nil, fmt.Errorf("connecting server: %w", err)
}
return c1, c2, func() {
c1.Close()
c2.Close()
}, nil
})
}
// tests that the idle memory overhead of a Conn blocked in a read is
// reasonable (under 2K). It was previously over 8KB with two 4KB
// buffers for rx/tx. This make sure we don't regress. Hopefully it
// doesn't turn into a flaky test. If so, const max can be adjusted,
// or it can be deleted or reworked.
func TestConnMemoryOverhead(t *testing.T) {
num := 1000
if testing.Short() {
num = 100
}
ng0 := runtime.NumGoroutine()
runtime.GC()
var ms0 runtime.MemStats
runtime.ReadMemStats(&ms0)
var closers []io.Closer
closeAll := func() {
for _, c := range closers {
c.Close()
}
closers = nil
}
defer closeAll()
for i := 0; i < num; i++ {
client, server := pair(t)
closers = append(closers, client, server)
go func() {
var buf [1]byte
client.Read(buf[:])
}()
}
t0 := time.Now()
deadline := t0.Add(3 * time.Second)
var ngo int
for time.Now().Before(deadline) {
runtime.GC()
ngo = runtime.NumGoroutine()
if ngo >= num {
break
}
time.Sleep(10 * time.Millisecond)
}
if ngo < num {
t.Fatalf("only %v goroutines; expected %v+", ngo, num)
}
runtime.GC()
var ms runtime.MemStats
runtime.ReadMemStats(&ms)
growthTotal := int64(ms.HeapAlloc) - int64(ms0.HeapAlloc)
growthEach := float64(growthTotal) / float64(num)
t.Logf("Alloced %v bytes, %.2f B/each", growthTotal, growthEach)
const max = 2000
if growthEach > max {
t.Errorf("allocated more than expected; want max %v bytes/each", max)
}
closeAll()
// And make sure our goroutines go away too.
deadline = time.Now().Add(3 * time.Second)
for time.Now().Before(deadline) {
ngo = runtime.NumGoroutine()
if ngo < ng0+num/10 {
break
}
time.Sleep(10 * time.Millisecond)
}
if ngo >= ng0+num/10 {
t.Errorf("goroutines didn't go back down; started at %v, now %v", ng0, ngo)
}
}
// mkConns creates synthetic Noise Conns wrapping the given net.Conns.
// This function is for testing just the Conn transport logic without
// having to muck about with Noise handshakes.
func mkConns(s1, s2 net.Conn) (*Conn, *Conn) {
var k1, k2 [chp.KeySize]byte
if _, err := rand.Read(k1[:]); err != nil {
panic(err)
}
if _, err := rand.Read(k2[:]); err != nil {
panic(err)
}
ret1 := &Conn{
conn: s1,
tx: txState{cipher: newCHP(k1)},
rx: rxState{cipher: newCHP(k2)},
}
ret2 := &Conn{
conn: s2,
tx: txState{cipher: newCHP(k2)},
rx: rxState{cipher: newCHP(k1)},
}
return ret1, ret2
}
type readSink struct {
r io.Reader
cond *sync.Cond
sync.Mutex
bs bytes.Buffer
err error
}
func sinkReads(r io.Reader) *readSink {
ret := &readSink{
r: r,
}
ret.cond = sync.NewCond(&ret.Mutex)
go func() {
var buf [4096]byte
for {
n, err := r.Read(buf[:])
ret.Lock()
ret.bs.Write(buf[:n])
if err != nil {
ret.err = err
}
ret.cond.Broadcast()
ret.Unlock()
if err != nil {
return
}
}
}()
return ret
}
func (s *readSink) String(total int) string {
s.Lock()
defer s.Unlock()
for s.bs.Len() < total && s.err == nil {
s.cond.Wait()
}
if s.err != nil {
total = s.bs.Len()
}
return string(s.bs.Bytes()[:total])
}
func (s *readSink) Error() error {
s.Lock()
defer s.Unlock()
for s.err == nil {
s.cond.Wait()
}
return s.err
}
func (s *readSink) Total() int {
s.Lock()
defer s.Unlock()
return s.bs.Len()
}
func pairWithConns(t *testing.T, clientConn, serverConn net.Conn) (*Conn, *Conn) {
var (
controlKey = key.NewMachine()
machineKey = key.NewMachine()
server *Conn
serverErr = make(chan error, 1)
)
go func() {
var err error
server, err = Server(context.Background(), serverConn, controlKey, testProtocolVersion, nil)
serverErr <- err
}()
client, err := Client(context.Background(), clientConn, machineKey, controlKey.Public(), testProtocolVersion)
if err != nil {
t.Fatalf("client connection failed: %v", err)
}
if err := <-serverErr; err != nil {
t.Fatalf("server connection failed: %v", err)
}
return client, server
}
func pair(t *testing.T) (*Conn, *Conn) {
s1, s2 := tsnettest.NewConn("noise", 128000)
return pairWithConns(t, s1, s2)
}