feat: agents connect by WebSocket, falling back to long polling and trying again each minute; auto, ws or lp per machine (0.1.0)

internal/agent/client.go

@@ -18,6 +18,7 @@ import (
 	"path/filepath"
 	"runtime"
 	"sync"
+	"sync/atomic"
 	"time"
 
 	lpra "github.com/x-grk/longpolling-remote-access"
@@ -53,6 +54,8 @@ type Agent struct {
 	pollCount int
 
 	// Direct tunnel currently running (see setTunnelAddr).
+	transport atomic.Value // string: auto, ws or lp, as the server says
+	link      atomic.Pointer[wsLink]
 	tunMu     sync.Mutex
 	tunAddr   string
 	tunCancel context.CancelFunc
@@ -107,10 +110,22 @@ func (a *Agent) Run(ctx context.Context) error {
 		a.setTunnelAddr(ctx, a.cfg.TunnelAddr)
 	}
 
+	// The control channel: a WebSocket, or long polls. In "auto" a failed or
+	// dropped WebSocket gives way to long polling for a minute, then is tried
+	// again; "ws" keeps trying the WebSocket; "lp" only polls.
 	backoff := time.Second
 	for ctx.Err() == nil {
-		if err := a.pollOnce(ctx); err != nil {
-			log.Printf("poll: %v (retry in %s)", err, backoff)
+		if a.mode() == proto.TransportLP {
+			a.longPoll(ctx, 0)
+			continue
+		}
+		err := a.runWS(ctx)
+		if ctx.Err() != nil {
+			return nil
+		}
+		switch a.mode() {
+		case proto.TransportWS:
+			log.Printf("websocket: %v (retry in %s)", err, backoff)
 			select {
 			case <-time.After(backoff):
 			case <-ctx.Done():
@@ -120,10 +135,71 @@ func (a *Agent) Run(ctx context.Context) error {
 				backoff *= 2
 			}
 			continue
+		case proto.TransportAuto:
+			log.Printf("websocket: %v - long polling, websocket again in a minute", err)
+		}
+		backoff = time.Second
+		a.longPoll(ctx, time.Minute)
+	}
+	return nil
+}
+
+// mode is the transport the server set for this agent.
+func (a *Agent) mode() string {
+	if v, ok := a.transport.Load().(string); ok {
+		return v
+	}
+	return proto.TransportAuto
+}
+
+func (a *Agent) applyTransport(v string) {
+	switch v {
+	case proto.TransportAuto, proto.TransportWS, proto.TransportLP:
+		if v != a.mode() {
+			log.Printf("transport: %s", v)
+			a.transport.Store(v)
+		}
+	}
+}
+
+// longPoll runs control polls: for d, or with d zero until the transport is
+// no longer long polling.
+func (a *Agent) longPoll(ctx context.Context, d time.Duration) {
+	until := time.Now().Add(d)
+	backoff := time.Second
+	for ctx.Err() == nil {
+		if d > 0 && (time.Now().After(until) || a.mode() == proto.TransportWS) {
+			return
+		}
+		if d == 0 && a.mode() != proto.TransportLP {
+			return
+		}
+		// In "auto" the minute is a minute: a poll still held then is cut
+		// short (the server keeps its job for the next one).
+		reqCtx, cancel := ctx, context.CancelFunc(func() {})
+		if d > 0 {
+			reqCtx, cancel = context.WithDeadline(ctx, until)
+		}
+		err := a.pollOnce(ctx, reqCtx)
+		cut := reqCtx.Err() == context.DeadlineExceeded
+		cancel()
+		if err != nil && cut && ctx.Err() == nil {
+			return
+		}
+		if err != nil {
+			log.Printf("poll: %v (retry in %s)", err, backoff)
+			select {
+			case <-time.After(backoff):
+			case <-ctx.Done():
+				return
+			}
+			if backoff < 15*time.Second {
+				backoff *= 2
+			}
+			continue
 		}
 		backoff = time.Second
 	}
-	return nil
 }
 
 func (a *Agent) enroll(ctx context.Context) error {
@@ -157,8 +233,10 @@ func (a *Agent) enroll(ctx context.Context) error {
 }
 
 // pollOnce performs one control long-poll and dispatches any instruction.
-func (a *Agent) pollOnce(ctx context.Context) error {
-	req, _ := http.NewRequestWithContext(ctx, http.MethodGet, a.cfg.Host+"/api/agent/poll", nil)
+// The poll itself runs on reqCtx, which may end early; what it starts runs on
+// ctx.
+func (a *Agent) pollOnce(ctx, reqCtx context.Context) error {
+	req, _ := http.NewRequestWithContext(reqCtx, http.MethodGet, a.cfg.Host+"/api/agent/poll", nil)
 	a.authHeaders(req)
 	// Report the version and our OS user occasionally (not every poll), so the
 	// panel can flag outdated agents without extra traffic.
@@ -183,7 +261,14 @@ func (a *Agent) pollOnce(ctx context.Context) error {
 		return err
 	}
 	a.applyTunnelHeader(ctx, resp.Header)
-	// Dispatch in the background so we keep polling for more work.
+	a.applyTransport(resp.Header.Get(proto.TransportHeader))
+	a.dispatch(ctx, msg)
+	return nil
+}
+
+// dispatch starts what a control message asks for, in the background, so the
+// control channel keeps listening.
+func (a *Agent) dispatch(ctx context.Context, msg proto.ControlMessage) {
 	switch msg.Type {
 	case proto.CtrlOpenTerminal:
 		go a.runTerminal(ctx, msg.SessionID, msg.User)
@@ -192,7 +277,6 @@ func (a *Agent) pollOnce(ctx context.Context) error {
 	case proto.CtrlTask:
 		go a.runTask(ctx, msg.TaskID)
 	}
-	return nil
 }
 
 // fetchConfig asks the server for the settings it hands to agents and reports
@@ -207,13 +291,17 @@ func (a *Agent) fetchConfig(ctx context.Context) bool {
 		return false
 	}
 	defer resp.Body.Close()
+	a.applyTransport(resp.Header.Get(proto.TransportHeader))
 	return a.applyTunnelHeader(ctx, resp.Header)
 }
 
 // applyTunnelHeader follows the tunnel address the server advertises as
 // "host:port"; an empty host means the host we already reach the server on.
 func (a *Agent) applyTunnelHeader(ctx context.Context, h http.Header) bool {
-	v := h.Get("X-Tunnel-Addr")
+	return a.applyTunnelAddr(ctx, h.Get("X-Tunnel-Addr"))
+}
+
+func (a *Agent) applyTunnelAddr(ctx context.Context, v string) bool {
 	if v == "" || a.cfg.TunnelPinned {
 		return false
 	}