fix: stabilize federated agent runtime updates (#550)

This commit is contained in:
sagit
2026-08-25 17:04:04 +08:00
committed by GitHub
parent 5041c722c9
commit f26014579b
9 changed files with 535 additions and 26 deletions
+117
View File
@@ -0,0 +1,117 @@
package socket
import (
"net"
"sync"
"testing"
"time"
coreservice "github.com/go-gost/core/service"
"github.com/go-gost/x/config"
"github.com/go-gost/x/registry"
)
type blockingCommandService struct {
started chan struct{}
release chan struct{}
startedOnce sync.Once
}
func (s *blockingCommandService) Serve() error { return nil }
func (s *blockingCommandService) Addr() net.Addr { return nil }
func (s *blockingCommandService) Close() error {
s.startedOnce.Do(func() { close(s.started) })
<-s.release
return nil
}
func TestMutationCommandsAreSerialized(t *testing.T) {
mutations := []string{
"AddService", "UpdateService", "DeleteService", "PauseService", "ResumeService",
"AddChains", "UpdateChains", "DeleteChains",
"AddLimiters", "UpdateLimiters", "DeleteLimiters",
"AddCLimiters", "UpdateCLimiters", "DeleteCLimiters",
"SetProtocol", "UpgradeAgent", "RollbackAgent", "reload",
}
for _, command := range mutations {
if !isMutationCommand(command) {
t.Fatalf("expected %s to use the serialized mutation queue", command)
}
}
}
func TestReadOnlyCommandsRemainBoundedAsync(t *testing.T) {
for _, command := range []string{"TcpPing", "UdpPing", "ServiceMonitorCheck"} {
if isMutationCommand(command) {
t.Fatalf("expected %s to remain a read-only command", command)
}
}
}
func TestCommandResponseTypePreservesRequestContract(t *testing.T) {
if got := commandResponseType("UpdateService"); got != "UpdateServiceResponse" {
t.Fatalf("unexpected response type: %s", got)
}
if got := commandResponseType(""); got != "UnknownCommandResponse" {
t.Fatalf("unexpected empty command response type: %s", got)
}
}
func TestMutationQueueExecutesCommandsInArrivalOrder(t *testing.T) {
originalConfig := config.Global()
t.Cleanup(func() { config.Set(originalConfig) })
firstName := "mutation_queue_first_tdd"
secondName := "mutation_queue_second_tdd"
first := &blockingCommandService{started: make(chan struct{}), release: make(chan struct{})}
second := &blockingCommandService{started: make(chan struct{}), release: make(chan struct{})}
for _, name := range []string{firstName, secondName} {
registry.ServiceRegistry().Unregister(name)
}
t.Cleanup(func() {
select {
case <-first.release:
default:
close(first.release)
}
select {
case <-second.release:
default:
close(second.release)
}
registry.ServiceRegistry().Unregister(firstName)
registry.ServiceRegistry().Unregister(secondName)
})
if err := registry.ServiceRegistry().Register(firstName, coreservice.Service(first)); err != nil {
t.Fatalf("register first service: %v", err)
}
if err := registry.ServiceRegistry().Register(secondName, coreservice.Service(second)); err != nil {
t.Fatalf("register second service: %v", err)
}
config.Set(&config.Config{Services: []*config.ServiceConfig{{Name: firstName}, {Name: secondName}}})
reporter := NewWebSocketReporter("", "mutation-queue-test-secret")
go reporter.runMutationCommands()
t.Cleanup(reporter.Stop)
reporter.dispatchCommand(CommandMessage{Type: "DeleteService", Data: map[string]any{"services": []string{firstName}}})
reporter.dispatchCommand(CommandMessage{Type: "DeleteService", Data: map[string]any{"services": []string{secondName}}})
select {
case <-first.started:
case <-time.After(time.Second):
t.Fatal("first mutation did not start")
}
select {
case <-second.started:
t.Fatal("second mutation started before first mutation completed")
case <-time.After(100 * time.Millisecond):
}
close(first.release)
select {
case <-second.started:
case <-time.After(time.Second):
t.Fatal("second mutation did not start after first mutation completed")
}
close(second.release)
}
+82 -18
View File
@@ -15,6 +15,11 @@ import (
xservice "github.com/go-gost/x/service"
)
type serviceReplacement struct {
config config.ServiceConfig
oldConfig *config.ServiceConfig
}
func createServices(req createServicesRequest) error {
if len(req.Data) == 0 {
@@ -95,11 +100,11 @@ func updateServices(req updateServicesRequest) error {
req.Data[i].Name = name
}
// 第二阶段:逐个更新服务(Upsert模式:存在则更新,不存在则创建)
changedServices := make([]struct {
config config.ServiceConfig
service service.Service
}, 0, len(req.Data))
// 第二阶段:逐个更新服务(Upsert模式:存在则更新,不存在则创建)。
// 配置变更命令由 WebSocket reporter 串行调度,但这里仍保留完整回滚,
// 避免新配置解析或监听失败后旧服务永久消失。
originalConfig := config.Global()
changedServices := make([]serviceReplacement, 0, len(req.Data))
for i := range req.Data {
serviceConfig := &req.Data[i]
name := serviceConfig.Name
@@ -107,33 +112,48 @@ func updateServices(req updateServicesRequest) error {
continue
}
// 1. 获取旧服务
old := registry.ServiceRegistry().Get(name)
var oldConfig *config.ServiceConfig
if originalConfig != nil {
for _, current := range originalConfig.Services {
if current != nil && strings.TrimSpace(current.Name) == name {
oldConfig = current
break
}
}
}
// 2. 关闭旧服务 (如果存在)
if old != nil {
// 3. 从注册表移除旧服务;registry 会负责关闭旧服务。
// 1. 关闭并移除旧服务(如果存在)。同名监听必须先释放端口,
// 才能创建新 listener。
if registry.ServiceRegistry().Get(name) != nil {
registry.ServiceRegistry().Unregister(name)
}
// 4. 解析新服务配置
// 2. 解析新服务配置。
svc, err := parser.ParseService(serviceConfig)
if err != nil {
rollbackErr := restoreServiceRuntime(name, oldConfig)
rollbackErr = errors.Join(rollbackErr, rollbackServiceReplacements(changedServices))
if rollbackErr != nil {
return fmt.Errorf("create service %s failed: %v; restore previous service failed: %w", name, err, rollbackErr)
}
return errors.New("create service " + name + " failed: " + err.Error())
}
changedServices = append(changedServices, struct {
config config.ServiceConfig
service service.Service
}{*serviceConfig, svc})
// 5. 注册新服务
// 3. 注册并启动新服务。
if err := registry.ServiceRegistry().Register(name, svc); err != nil {
svc.Close()
rollbackErr := restoreServiceRuntime(name, oldConfig)
rollbackErr = errors.Join(rollbackErr, rollbackServiceReplacements(changedServices))
if rollbackErr != nil {
return fmt.Errorf("service %s already exists; restore previous service failed: %w", name, rollbackErr)
}
return errors.New("service " + name + " already exists")
}
// 6. 启动新服务
go svc.Serve()
changedServices = append(changedServices, serviceReplacement{
config: *serviceConfig,
oldConfig: oldConfig,
})
}
if len(changedServices) == 0 {
return nil
@@ -158,12 +178,56 @@ func updateServices(req updateServicesRequest) error {
}
return nil
}); err != nil {
config.Set(originalConfig)
if rollbackErr := rollbackServiceReplacements(changedServices); rollbackErr != nil {
return fmt.Errorf("%w; restore previous services failed: %v", err, rollbackErr)
}
return err
}
return nil
}
func restoreServiceRuntime(name string, serviceConfig *config.ServiceConfig) error {
name = strings.TrimSpace(name)
if name == "" || serviceConfig == nil {
return nil
}
if registry.ServiceRegistry().Get(name) != nil {
registry.ServiceRegistry().Unregister(name)
}
cfgCopy := *serviceConfig
cfgCopy.Name = name
svc, err := parser.ParseService(&cfgCopy)
if err != nil {
return err
}
if err := registry.ServiceRegistry().Register(name, svc); err != nil {
svc.Close()
return err
}
go svc.Serve()
return nil
}
func rollbackServiceReplacements(replacements []serviceReplacement) error {
var rollbackErr error
for i := len(replacements) - 1; i >= 0; i-- {
name := strings.TrimSpace(replacements[i].config.Name)
if name == "" {
continue
}
if registry.ServiceRegistry().Get(name) != nil {
registry.ServiceRegistry().Unregister(name)
}
if err := restoreServiceRuntime(name, replacements[i].oldConfig); err != nil {
rollbackErr = errors.Join(rollbackErr, fmt.Errorf("restore service %s: %w", name, err))
}
}
return rollbackErr
}
func serviceConfigUnchanged(name string, next config.ServiceConfig) bool {
cfg := config.Global()
if cfg == nil {
+39
View File
@@ -7,6 +7,8 @@ import (
corelogger "github.com/go-gost/core/logger"
"github.com/go-gost/core/service"
"github.com/go-gost/x/config"
_ "github.com/go-gost/x/handler/auto"
_ "github.com/go-gost/x/listener/tcp"
xlogger "github.com/go-gost/x/logger"
"github.com/go-gost/x/registry"
)
@@ -15,6 +17,43 @@ type recordingService struct {
closed int
}
func TestUpdateServicesParseFailureRestoresPreviousRuntime(t *testing.T) {
corelogger.SetDefault(xlogger.Nop())
name := "restore_after_failed_update_tdd"
existing := &recordingService{}
registry.ServiceRegistry().Unregister(name)
t.Cleanup(func() { registry.ServiceRegistry().Unregister(name) })
if err := registry.ServiceRegistry().Register(name, service.Service(existing)); err != nil {
t.Fatalf("register existing service: %v", err)
}
originalConfig := config.Global()
t.Cleanup(func() { config.Set(originalConfig) })
serviceConfig := config.ServiceConfig{Name: name, Addr: "127.0.0.1:0"}
config.Set(&config.Config{Services: []*config.ServiceConfig{&serviceConfig}})
invalid := serviceConfig
invalid.Listener = &config.ListenerConfig{Type: "listener-does-not-exist"}
if err := updateServices(updateServicesRequest{Data: []config.ServiceConfig{invalid}}); err == nil {
t.Fatalf("expected invalid service update to fail")
}
if existing.closed != 1 {
t.Fatalf("expected old runtime to be closed once, got %d", existing.closed)
}
if registry.ServiceRegistry().Get(name) == nil {
t.Fatalf("expected previous runtime to be restored")
}
cfg := config.Global()
if len(cfg.Services) != 1 || cfg.Services[0] == nil || cfg.Services[0].Name != name {
t.Fatalf("expected previous config to remain, got %#v", cfg.Services)
}
if cfg.Services[0].Listener != nil {
t.Fatalf("expected previous listener config to remain unchanged")
}
}
func (s *recordingService) Serve() error { return nil }
func (s *recordingService) Addr() net.Addr { return nil }
func (s *recordingService) Close() error {
+90 -4
View File
@@ -16,6 +16,7 @@ import (
"os"
"os/exec"
"runtime"
"runtime/debug"
"strconv"
"strings"
"sync"
@@ -152,6 +153,8 @@ const (
maxBackoff = 2 * time.Minute // 重连最大退避
defaultMetricReportInterval = 5 * time.Second
maxConcurrentTCPPings = 8
maxConcurrentReadCommands = 16
maxQueuedMutationCommands = 256
)
type WebSocketReporter struct {
@@ -174,6 +177,8 @@ type WebSocketReporter struct {
connMutex sync.Mutex // 连接状态锁
aesCrypto *crypto.AESCrypto // AES加密器
tcpPingSem chan struct{} // 限制诊断探测并发,避免离线目标耗尽连接
readCommandSem chan struct{} // 限制只读命令并发,避免诊断请求耗尽资源
mutationQueue chan CommandMessage
}
var wsDial = func(dialer *websocket.Dialer, rawURL string) (*websocket.Conn, *http.Response, error) {
@@ -204,6 +209,8 @@ func NewWebSocketReporter(serverURL string, secret string) *WebSocketReporter {
connecting: false,
aesCrypto: aesCrypto,
tcpPingSem: make(chan struct{}, maxConcurrentTCPPings),
readCommandSem: make(chan struct{}, maxConcurrentReadCommands),
mutationQueue: make(chan CommandMessage, maxQueuedMutationCommands),
}
}
@@ -231,6 +238,7 @@ func (w *WebSocketReporter) releaseTCPPingSlot() {
// Start 启动WebSocket报告器
func (w *WebSocketReporter) Start() {
go w.runMutationCommands()
go w.run()
}
@@ -777,8 +785,7 @@ func (w *WebSocketReporter) handleReceivedMessage(messageType int, message []byt
}
if cmdMsg.Type != "call" {
// 所有命令统一异步执行,避免阻塞消息接收循环
go w.routeCommand(cmdMsg)
w.dispatchCommand(cmdMsg)
}
} else {
// 处理普通消息
@@ -789,8 +796,7 @@ func (w *WebSocketReporter) handleReceivedMessage(messageType int, message []byt
return
}
if cmdMsg.Type != "call" {
// 所有命令统一异步执行,避免阻塞消息接收循环
go w.routeCommand(cmdMsg)
w.dispatchCommand(cmdMsg)
}
}
@@ -799,6 +805,86 @@ func (w *WebSocketReporter) handleReceivedMessage(messageType int, message []byt
}
}
// dispatchCommand keeps all runtime mutations ordered while allowing bounded
// concurrency for read-only diagnostics. Mutations share process-wide
// registries and configuration, so running them concurrently can corrupt the
// persisted config or interleave service lifecycle operations.
func (w *WebSocketReporter) dispatchCommand(cmd CommandMessage) {
if isMutationCommand(cmd.Type) {
select {
case w.mutationQueue <- cmd:
case <-w.ctx.Done():
w.sendCommandFailure(cmd, "Agent is shutting down")
default:
w.sendCommandFailure(cmd, "运行时配置命令队列已满,请稍后重试")
}
return
}
select {
case w.readCommandSem <- struct{}{}:
go func() {
defer func() { <-w.readCommandSem }()
w.routeCommandSafely(cmd)
}()
case <-w.ctx.Done():
w.sendCommandFailure(cmd, "Agent is shutting down")
default:
w.sendCommandFailure(cmd, "只读命令并发过多,请稍后重试")
}
}
func (w *WebSocketReporter) runMutationCommands() {
for {
select {
case <-w.ctx.Done():
return
case cmd := <-w.mutationQueue:
w.routeCommandSafely(cmd)
}
}
}
func (w *WebSocketReporter) routeCommandSafely(cmd CommandMessage) {
defer func() {
if recovered := recover(); recovered != nil {
fmt.Printf("❌ 命令处理 panic: type=%s panic=%v\n%s", cmd.Type, recovered, debug.Stack())
w.sendCommandFailure(cmd, fmt.Sprintf("命令处理异常: %v", recovered))
}
}()
w.routeCommand(cmd)
}
func (w *WebSocketReporter) sendCommandFailure(cmd CommandMessage, message string) {
w.sendResponse(CommandResponse{
Type: commandResponseType(cmd.Type),
Success: false,
Message: message,
RequestId: cmd.RequestId,
})
}
func commandResponseType(commandType string) string {
commandType = strings.TrimSpace(commandType)
if commandType == "" {
return "UnknownCommandResponse"
}
return commandType + "Response"
}
func isMutationCommand(commandType string) bool {
switch strings.ToLower(strings.TrimSpace(commandType)) {
case "addservice", "updateservice", "deleteservice", "pauseservice", "resumeservice",
"addchains", "updatechains", "deletechains",
"addlimiters", "updatelimiters", "deletelimiters",
"addclimiters", "updateclimiters", "deleteclimiters",
"setprotocol", "upgradeagent", "rollbackagent", "reload":
return true
default:
return false
}
}
// routeCommand 路由命令到对应的处理函数
func (w *WebSocketReporter) routeCommand(cmd CommandMessage) {
jsonBytes, errs := json.Marshal(cmd)