fix: synchronize service startup and shutdown (#5560)

This commit is contained in:
fatedier
2026-10-09 12:27:12 +08:00
committed by GitHub
parent 4cd4747525
commit 55107d003d
6 changed files with 485 additions and 14 deletions
+2 -1
View File
@@ -267,8 +267,9 @@ func (pw *Wrapper) InWorkConn(workConn net.Conn, m *msg.StartWorkConn) {
xl := pw.xl
pw.mu.RLock()
pxy := pw.pxy
running := pw.Phase == ProxyPhaseRunning
pw.mu.RUnlock()
if pxy != nil && pw.Phase == ProxyPhaseRunning {
if pxy != nil && running {
xl.Debugf("start a new work connection, localAddr: %s remoteAddr: %s", workConn.LocalAddr().String(), workConn.RemoteAddr().String())
go pxy.InWorkConn(workConn, m)
} else {
+76
View File
@@ -0,0 +1,76 @@
//go:build !frps
package proxy
import (
"io"
"net"
"sync/atomic"
"testing"
"time"
"github.com/stretchr/testify/require"
"github.com/fatedier/frp/pkg/msg"
"github.com/fatedier/frp/pkg/util/xlog"
)
type closingWorkProxy struct {
Proxy
accepted atomic.Int64
}
func (p *closingWorkProxy) InWorkConn(conn net.Conn, _ *msg.StartWorkConn) {
p.accepted.Add(1)
_ = conn.Close()
}
func assertWrapperWorkConnClosed(t *testing.T, pw *Wrapper) {
t.Helper()
conn, peer := net.Pipe()
defer conn.Close()
defer peer.Close()
require.NoError(t, peer.SetReadDeadline(time.Now().Add(time.Second)))
pw.InWorkConn(conn, &msg.StartWorkConn{})
_, err := peer.Read(make([]byte, 1))
require.ErrorIs(t, err, io.EOF)
}
func TestWrapperInWorkConnPhaseSnapshot(t *testing.T) {
pxy := &closingWorkProxy{}
pw := &Wrapper{pxy: pxy, xl: xlog.New()}
for _, phase := range []string{ProxyPhaseNew, ProxyPhaseWaitStart, ProxyPhaseStartErr, ProxyPhaseCheckFailed, ProxyPhaseClosed} {
pw.Phase = phase
assertWrapperWorkConnClosed(t, pw)
}
require.Zero(t, pxy.accepted.Load(), "inactive proxy received a work connection")
pw.Phase = ProxyPhaseRunning
pw.pxy = nil
assertWrapperWorkConnClosed(t, pw)
pw.pxy = pxy
assertWrapperWorkConnClosed(t, pw)
require.Equal(t, int64(1), pxy.accepted.Load())
start, done := make(chan struct{}), make(chan struct{})
go func() {
<-start
for range 20000 {
pw.mu.Lock()
pw.Phase = ProxyPhaseRunning
pw.mu.Unlock()
pw.mu.Lock()
pw.Phase = ProxyPhaseClosed
pw.mu.Unlock()
}
close(done)
}()
close(start)
for range 10000 {
assertWrapperWorkConnClosed(t, pw)
}
select {
case <-done:
case <-time.After(3 * time.Second):
t.Fatal("phase writer did not finish")
}
}
+29 -4
View File
@@ -155,6 +155,9 @@ type Service struct {
ctx context.Context
// call cancel to stop service
cancel context.CancelCauseFunc
// Protect cancel publication and shutdown before the first Run.
lifecycleMu sync.Mutex
closed bool
connectorCreator func(context.Context, *v1.ClientCommonConfig) Connector
handleWorkConnCb func(*v1.ProxyBaseConfig, net.Conn, *msg.StartWorkConn) bool
@@ -225,8 +228,20 @@ func NewService(options ServiceOptions) (*Service, error) {
func (svr *Service) Run(ctx context.Context) error {
ctx, cancel := context.WithCancelCause(ctx)
svr.lifecycleMu.Lock()
if svr.closed {
svr.lifecycleMu.Unlock()
cancel(nil)
return ctx.Err()
}
svr.ctx = xlog.NewContext(ctx, xlog.FromContextSafe(ctx))
svr.cancel = cancel
svr.lifecycleMu.Unlock()
defer cancel(nil)
defer svr.stop()
if err := ctx.Err(); err != nil {
return err
}
// set custom DNSServer
if svr.common.DNSServer != "" {
@@ -237,7 +252,6 @@ func (svr *Service) Run(ctx context.Context) error {
vnetController := svr.vnetController
if err := svr.vnetController.Init(); err != nil {
log.Errorf("init virtual network controller error: %v", err)
svr.stop()
return err
}
go func() {
@@ -263,14 +277,12 @@ func (svr *Service) Run(ctx context.Context) error {
if svr.ctl == nil {
cancelCause := cancelErr{}
_ = errors.As(context.Cause(svr.ctx), &cancelCause)
svr.stop()
return fmt.Errorf("login to the server failed: %v. With loginFailExit enabled, no additional retries will be attempted", cancelCause.Err)
}
go svr.keepControllerWorking()
<-svr.ctx.Done()
svr.stop()
return nil
}
@@ -428,7 +440,20 @@ func (svr *Service) Close() {
func (svr *Service) GracefulClose(d time.Duration) {
svr.gracefulShutdownDuration.Store(int64(d))
svr.cancel(nil)
svr.lifecycleMu.Lock()
if svr.closed {
svr.lifecycleMu.Unlock()
return
}
svr.closed = true
cancel := svr.cancel
svr.lifecycleMu.Unlock()
if cancel != nil {
// Once Run has claimed startup, it also owns cleanup.
cancel(nil)
} else {
svr.stop()
}
}
func (svr *Service) stop() {
+186
View File
@@ -0,0 +1,186 @@
package client
import (
"context"
"io"
"net"
"sync/atomic"
"testing"
"time"
"github.com/stretchr/testify/require"
"github.com/fatedier/frp/pkg/config/source"
v1 "github.com/fatedier/frp/pkg/config/v1"
"github.com/fatedier/frp/pkg/config/v1/validation"
"github.com/fatedier/frp/pkg/msg"
)
func newLifecycleTestService(t *testing.T, creator func(context.Context, *v1.ClientCommonConfig) Connector) (*Service, string) {
t.Helper()
svr, err := NewService(ServiceOptions{
Common: &v1.ClientCommonConfig{
LoginFailExit: new(false),
WebServer: v1.WebServerConfig{Addr: "127.0.0.1", Port: getFreeTCPPort(t)},
},
ConfigSourceAggregator: source.NewAggregator(source.NewConfigSource()),
ConnectorCreator: creator,
})
require.NoError(t, err)
_, err = validation.NewConfigValidator(nil).ValidateClientCommonConfig(svr.common)
require.NoError(t, err)
address := svr.webServer.Address()
assertLifecycleAdminPort(t, address, false)
return svr, address
}
func assertLifecycleAdminPort(t *testing.T, address string, released bool) {
t.Helper()
ln, err := net.Listen("tcp", address)
if ln != nil {
_ = ln.Close()
}
if released {
require.NoError(t, err, "admin listener was not released")
} else {
require.Error(t, err, "admin listener was not bound")
}
}
func waitLifecycleSignal(t *testing.T, done <-chan struct{}) {
t.Helper()
select {
case <-done:
case <-time.After(3 * time.Second):
t.Fatal("service lifecycle operation did not finish")
}
}
func TestServiceRunAfterEarlyShutdown(t *testing.T) {
for _, mode := range []string{"close before Run", "parent already cancelled"} {
t.Run(mode, func(t *testing.T) {
var attempts atomic.Int64
svr, address := newLifecycleTestService(t, func(context.Context, *v1.ClientCommonConfig) Connector {
attempts.Add(1)
return &failingConnector{err: net.ErrClosed}
})
// Detect startup's global DNS side effect without making DNS requests.
resolver := net.DefaultResolver
t.Cleanup(func() { net.DefaultResolver = resolver })
svr.common.DNSServer = "127.0.0.1:53"
ctx, cancel := context.WithCancel(context.Background())
defer cancel()
if mode == "close before Run" {
svr.Close()
assertLifecycleAdminPort(t, address, true)
} else {
cancel()
}
done := make(chan struct{})
var runErr error
go func() {
runErr = svr.Run(ctx)
close(done)
}()
waitLifecycleSignal(t, done)
require.ErrorIs(t, runErr, context.Canceled)
require.Zero(t, attempts.Load(), "closed service attempted login")
require.Same(t, resolver, net.DefaultResolver, "closed service changed DNS")
assertLifecycleAdminPort(t, address, true)
})
}
}
func TestServiceStartupShutdownLeavesCleanupToRun(t *testing.T) {
for _, mode := range []string{"Close", "parent cancellation"} {
t.Run(mode, func(t *testing.T) {
entered, resume := make(chan struct{}), make(chan struct{})
svr, address := newLifecycleTestService(t, func(ctx context.Context, _ *v1.ClientCommonConfig) Connector {
close(entered)
<-resume
return &failingConnector{err: ctx.Err()}
})
ctx, cancel := context.WithCancel(context.Background())
defer cancel()
defer close(resume)
runDone := make(chan struct{})
go func() { _ = svr.Run(ctx); close(runDone) }()
waitLifecycleSignal(t, entered)
closeDone := make(chan struct{})
go func() {
if mode == "Close" {
svr.Close()
} else {
cancel()
}
close(closeDone)
}()
waitLifecycleSignal(t, closeDone)
select {
case <-runDone:
t.Fatal("Run exited while startup still owned its resources")
default:
}
assertLifecycleAdminPort(t, address, false)
resume <- struct{}{}
waitLifecycleSignal(t, runDone)
assertLifecycleAdminPort(t, address, true)
require.ErrorIs(t, svr.ctx.Err(), context.Canceled)
})
}
}
func TestServiceShutdownAfterLogin(t *testing.T) {
for _, mode := range []string{"Close", "parent cancellation"} {
t.Run(mode, func(t *testing.T) {
clientConn, serverConn := net.Pipe()
defer clientConn.Close()
defer serverConn.Close()
connector := &testConnector{conn: clientConn}
svr, address := newLifecycleTestService(t, func(context.Context, *v1.ClientCommonConfig) Connector {
return connector
})
defer svr.Close()
peerDone := make(chan error, 1)
go func() {
rw := msg.NewV1ReadWriter(serverConn)
var login msg.Login
if err := rw.ReadMsgInto(&login); err != nil {
peerDone <- err
return
}
if err := rw.WriteMsg(&msg.LoginResp{RunID: "lifecycle-test"}); err != nil {
peerDone <- err
return
}
_, err := io.Copy(io.Discard, serverConn)
peerDone <- err
}()
ctx, cancel := context.WithCancel(context.Background())
defer cancel()
runDone := make(chan struct{})
var runErr error
go func() { runErr = svr.Run(ctx); close(runDone) }()
require.Eventually(t, func() bool {
svr.ctlMu.RLock()
defer svr.ctlMu.RUnlock()
return svr.ctl != nil
}, 3*time.Second, time.Millisecond, "login did not establish control")
if mode == "Close" {
svr.Close()
} else {
cancel()
}
waitLifecycleSignal(t, runDone)
require.NoError(t, runErr)
require.True(t, connector.closed.Load())
select {
case err := <-peerDone:
require.NoError(t, err)
case <-time.After(3 * time.Second):
t.Fatal("control connection was not released")
}
assertLifecycleAdminPort(t, address, true)
})
}
}