fix(rpc): cap concurrent IO streams per user and per server

GHSA-jg62-j5h6-8mpq: the terminal and file-manager endpoints created unbounded
IO streams; an authenticated member could open thousands, each spawning
goroutines, a 1MiB buffer and an agent-side PTY, exhausting dashboard and agent
resources. CreateStream now enforces a per-user (20) and per-server (40) cap in
the existing ioStreamMutex critical section, using the stream map as the single
source of truth. Dashboard-internal streams (uid==0: NAT, server transfer, MCP
transfer) skip the per-user cap but still count per-server. Adds caps,
exemption, slot-release and no-leak regression tests.

cmd/dashboard/controller/fm.go

@@ -49,7 +49,9 @@ func createFM(c *gin.Context) (*model.CreateFMResponse, error) {
 		return nil, err
 	}
 
-	rpc.NezhaHandlerSingleton.CreateStream(streamId, getUid(c), server.ID)
+	if err := rpc.NezhaHandlerSingleton.CreateStream(streamId, getUid(c), server.ID); err != nil {
+		return nil, err
+	}
 
 	fmData, _ := json.Marshal(&model.TaskFM{
 		StreamID: streamId,

cmd/dashboard/controller/mcp_transfer.go

@@ -962,7 +962,9 @@ func openFsTransferStream(ctx context.Context, serverID uint64, req *model.FsTra
 	}
 	req.StreamID = streamId
 
-	rpc.NezhaHandlerSingleton.CreateStreamWithPurpose(streamId, 0, serverID, rpc.PurposeMCPTransfer)
+	if err := rpc.NezhaHandlerSingleton.CreateStreamWithPurpose(streamId, 0, serverID, rpc.PurposeMCPTransfer); err != nil {
+		return nil, func() {}, err
+	}
 	cleanup := func() { _ = rpc.NezhaHandlerSingleton.CloseStream(streamId) }
 
 	body, err := json.Marshal(req)

cmd/dashboard/controller/terminal.go

@@ -47,7 +47,9 @@ func createTerminal(c *gin.Context) (*model.CreateTerminalResponse, error) {
 		return nil, err
 	}
 
-	rpc.NezhaHandlerSingleton.CreateStream(streamId, getUid(c), server.ID)
+	if err := rpc.NezhaHandlerSingleton.CreateStream(streamId, getUid(c), server.ID); err != nil {
+		return nil, err
+	}
 
 	terminalData, _ := json.Marshal(&model.TerminalTask{
 		StreamID: streamId,

cmd/dashboard/rpc/rpc.go

@@ -217,7 +217,11 @@ func ServeNAT(w http.ResponseWriter, r *http.Request, natConfig *model.NAT) {
 	// IS required though — the receiving agent must prove it is the server the
 	// NAT config addressed, otherwise any agent that snoops the streamId can
 	// answer NAT traffic on behalf of an unrelated host.
-	rpcService.NezhaHandlerSingleton.CreateStream(streamId, 0, server.ID)
+	if err := rpcService.NezhaHandlerSingleton.CreateStream(streamId, 0, server.ID); err != nil {
+		w.WriteHeader(http.StatusTooManyRequests)
+		w.Write(fmt.Appendf(nil, "stream limit: %v", err))
+		return
+	}
 	defer rpcService.NezhaHandlerSingleton.CloseStream(streamId)
 
 	taskData, err := json.Marshal(model.TaskNAT{

service/rpc/io_stream.go

@@ -48,14 +48,44 @@ var bufPool = sync.Pool{
 	},
 }
 
-func (s *NezhaHandler) CreateStream(streamId string, creatorUserID uint64, targetServerID uint64) {
-	s.CreateStreamWithPurpose(streamId, creatorUserID, targetServerID, PurposeLegacy)
+const (
+	maxStreamsPerUser   = 20
+	maxStreamsPerServer = 40
+)
+
+var (
+	ErrTooManyStreamsForUser   = errors.New("too many concurrent streams for this user")
+	ErrTooManyStreamsForServer = errors.New("too many concurrent streams for this server")
+)
+
+func (s *NezhaHandler) CreateStream(streamId string, creatorUserID uint64, targetServerID uint64) error {
+	return s.CreateStreamWithPurpose(streamId, creatorUserID, targetServerID, PurposeLegacy)
 }
 
-func (s *NezhaHandler) CreateStreamWithPurpose(streamId string, creatorUserID uint64, targetServerID uint64, purpose StreamPurpose) {
+func (s *NezhaHandler) CreateStreamWithPurpose(streamId string, creatorUserID uint64, targetServerID uint64, purpose StreamPurpose) error {
 	s.ioStreamMutex.Lock()
 	defer s.ioStreamMutex.Unlock()
 
+	var perUser, perServer int
+	for _, ctx := range s.ioStreams {
+		if creatorUserID != 0 && ctx.creatorUserID == creatorUserID {
+			perUser++
+		}
+		if ctx.targetServerID == targetServerID {
+			perServer++
+		}
+	}
+	// creatorUserID==0 is a dashboard-internal stream (NAT, server transfer,
+	// MCP transfer); only end-user-initiated streams are capped per user, but
+	// every stream counts toward the per-server cap so one server cannot be
+	// flooded regardless of who opened the streams.
+	if creatorUserID != 0 && perUser >= maxStreamsPerUser {
+		return ErrTooManyStreamsForUser
+	}
+	if perServer >= maxStreamsPerServer {
+		return ErrTooManyStreamsForServer
+	}
+
 	s.ioStreams[streamId] = &ioStreamContext{
 		creatorUserID:    creatorUserID,
 		targetServerID:   targetServerID,
@@ -64,6 +94,7 @@ func (s *NezhaHandler) CreateStreamWithPurpose(streamId string, creatorUserID ui
 		agentIoConnectCh: make(chan struct{}),
 		revokedCh:        make(chan struct{}),
 	}
+	return nil
 }
 
 // IsStreamAuthorizedForAgent reports whether the connecting agent is the
@@ -270,8 +301,6 @@ func (s *NezhaHandler) CloseStream(streamId string) error {
 	return nil
 }
 
-
-
 // UserConnected publishes the user-side IO under ioStreamMutex so concurrent
 // Revoke* / WaitForAgent / StartStream see a consistent stream view.
 // Without the lock, the bare assignment to stream.userIo races with the

service/rpc/io_stream_leak_test.go

@@ -0,0 +1,67 @@
+package rpc
+
+import (
+	"runtime"
+	"testing"
+	"time"
+)
+
+// settleGoroutines lets transient goroutines wind down so the count reflects
+// only durable leaks, not in-flight teardown.
+func settleGoroutines() int {
+	var n int
+	for i := 0; i < 50; i++ {
+		runtime.GC()
+		time.Sleep(20 * time.Millisecond)
+		n = runtime.NumGoroutine()
+	}
+	return n
+}
+
+// TestStartStream_NoGoroutineLeakAfterClose verifies the bidirectional relay in
+// StartStream does not strand a goroutine. StartStream launches two
+// io.CopyBuffer goroutines (user<-agent and agent<-user) but returns after the
+// first one finishes. The second goroutine stays blocked in CopyBuffer until
+// its endpoints are closed. CloseStream closes both endpoints, which must
+// unblock and drain that second goroutine. If it doesn't, every terminal / fm /
+// NAT session leaks one goroutine for the lifetime of the dashboard.
+func TestStartStream_NoGoroutineLeakAfterClose(t *testing.T) {
+	base := settleGoroutines()
+
+	const n = 20
+	for i := 0; i < n; i++ {
+		h := NewNezhaHandler()
+		const id = "leak-stream"
+
+		if err := h.CreateStream(id, 1, 1); err != nil {
+			t.Fatalf("CreateStream: %v", err)
+		}
+
+		userIo, agentIo := newPipeReadWriter(), newPipeReadWriter()
+		h.AgentConnected(id, agentIo)
+		h.UserConnected(id, userIo)
+
+		done := make(chan struct{})
+		go func() {
+			_ = h.StartStream(id, time.Second*5)
+			close(done)
+		}()
+
+		// Close one endpoint so the first CopyBuffer returns and StartStream
+		// unblocks, mirroring a peer disconnect.
+		time.Sleep(10 * time.Millisecond)
+		userIo.Close()
+		<-done
+
+		// The caller's defer CloseStream closes both endpoints, which must
+		// drain the still-blocked second copy goroutine.
+		_ = h.CloseStream(id)
+		agentIo.Close()
+	}
+
+	after := settleGoroutines()
+	if grew := after - base; grew > 2 {
+		t.Fatalf("goroutine leak in StartStream relay: ran %d streams, goroutines grew by %d (base=%d after=%d)",
+			n, grew, base, after)
+	}
+}

service/rpc/io_stream_test.go

@@ -1,6 +1,8 @@
 package rpc
 
 import (
+	"errors"
+	"fmt"
 	"io"
 	"reflect"
 	"testing"
@@ -98,6 +100,130 @@ func TestIOStream(t *testing.T) {
 	})
 }
 
+// The WebSocket stream endpoints (terminal / fm) were unbounded: an
+// authenticated member could open thousands of streams, each spawning
+// goroutines, a 1 MiB buffer, and an agent-side PTY, exhausting dashboard and
+// agent resources (GHSA-jg62-j5h6-8mpq). CreateStream now caps concurrent
+// streams per user and per server. These tests pin the caps and the
+// dashboard-internal (uid==0) exemption.
+
+// Baseline: a normal operator opening a terminal and a file-manager session
+// against one server (the everyday case) must always succeed — the cap exists
+// to stop floods, not to interfere with ordinary use.
+func TestCreateStreamNormalUserEverydayUseSucceeds(t *testing.T) {
+	h := NewNezhaHandler()
+	const uid, serverID = uint64(7), uint64(1)
+
+	if err := h.CreateStream("term", uid, serverID); err != nil {
+		t.Fatalf("opening a terminal must succeed for a normal user, got %v", err)
+	}
+	if err := h.CreateStream("fm", uid, serverID); err != nil {
+		t.Fatalf("opening a file manager alongside a terminal must succeed, got %v", err)
+	}
+}
+
+// Several normal users working at the same time must not interfere: one user's
+// streams do not consume another user's per-user budget.
+func TestCreateStreamNormalUsersAreIndependent(t *testing.T) {
+	h := NewNezhaHandler()
+
+	for u := uint64(1); u <= 5; u++ {
+		for i := 0; i < maxStreamsPerUser; i++ {
+			id := fmt.Sprintf("u%d-s%d", u, i)
+			if err := h.CreateStream(id, u, 100+u); err != nil {
+				t.Fatalf("user %d stream %d must succeed; per-user budgets must be independent, got %v", u, i, err)
+			}
+		}
+	}
+}
+
+func TestCreateStreamEnforcesPerUserCap(t *testing.T) {
+	h := NewNezhaHandler()
+	const uid = uint64(42)
+
+	for i := 0; i < maxStreamsPerUser; i++ {
+		if err := h.CreateStream(fmt.Sprintf("u-%d", i), uid, uint64(i)); err != nil {
+			t.Fatalf("stream %d within the per-user cap must succeed, got %v", i, err)
+		}
+	}
+
+	err := h.CreateStream("u-over", uid, 9999)
+	if !errors.Is(err, ErrTooManyStreamsForUser) {
+		t.Fatalf("the (maxStreamsPerUser+1)-th stream must be rejected with ErrTooManyStreamsForUser, got %v", err)
+	}
+}
+
+func TestCreateStreamEnforcesPerServerCap(t *testing.T) {
+	h := NewNezhaHandler()
+	const serverID = uint64(7)
+
+	for i := 0; i < maxStreamsPerServer; i++ {
+		if err := h.CreateStream(fmt.Sprintf("s-%d", i), uint64(i+1), serverID); err != nil {
+			t.Fatalf("stream %d within the per-server cap must succeed, got %v", i, err)
+		}
+	}
+
+	err := h.CreateStream("s-over", 99999, serverID)
+	if !errors.Is(err, ErrTooManyStreamsForServer) {
+		t.Fatalf("the (maxStreamsPerServer+1)-th stream to one server must be rejected with ErrTooManyStreamsForServer, got %v", err)
+	}
+}
+
+// Dashboard-internal streams (NAT, server transfer, MCP transfer) pass
+// creatorUserID==0. They must NOT be capped per user, or those features would
+// throttle themselves; but they must still count toward the per-server cap so
+// no single server can be flooded regardless of the originating path.
+func TestCreateStreamExemptsInternalStreamsFromPerUserCap(t *testing.T) {
+	h := NewNezhaHandler()
+
+	for i := 0; i < maxStreamsPerUser*3; i++ {
+		if err := h.CreateStream(fmt.Sprintf("internal-%d", i), 0, uint64(i)); err != nil {
+			t.Fatalf("internal stream %d (uid==0) must never hit the per-user cap, got %v", i, err)
+		}
+	}
+}
+
+func TestCreateStreamInternalStreamsStillCountTowardPerServerCap(t *testing.T) {
+	h := NewNezhaHandler()
+	const serverID = uint64(3)
+
+	for i := 0; i < maxStreamsPerServer; i++ {
+		if err := h.CreateStream(fmt.Sprintf("internal-s-%d", i), 0, serverID); err != nil {
+			t.Fatalf("internal stream %d within the per-server cap must succeed, got %v", i, err)
+		}
+	}
+
+	err := h.CreateStream("internal-s-over", 0, serverID)
+	if !errors.Is(err, ErrTooManyStreamsForServer) {
+		t.Fatalf("internal streams must still be subject to the per-server cap, got %v", err)
+	}
+}
+
+// Closing a stream must free its slot so a user who hit the cap can open new
+// streams after old ones end — otherwise normal churn would permanently lock
+// a user out.
+func TestCreateStreamFreesSlotAfterClose(t *testing.T) {
+	h := NewNezhaHandler()
+	const uid = uint64(55)
+
+	for i := 0; i < maxStreamsPerUser; i++ {
+		if err := h.CreateStream(fmt.Sprintf("c-%d", i), uid, 1); err != nil {
+			t.Fatalf("setup stream %d must succeed, got %v", i, err)
+		}
+	}
+	if err := h.CreateStream("c-over", uid, 1); !errors.Is(err, ErrTooManyStreamsForUser) {
+		t.Fatalf("expected per-user cap to be hit, got %v", err)
+	}
+
+	if err := h.CloseStream("c-0"); err != nil {
+		t.Fatalf("CloseStream failed: %v", err)
+	}
+
+	if err := h.CreateStream("c-after-close", uid, 1); err != nil {
+		t.Fatalf("after closing one stream the user must be able to open another, got %v", err)
+	}
+}
+
 func newPipeReadWriter() io.ReadWriteCloser {
 	r, w := io.Pipe()
 	return struct {