perf: reduce latency in 60 Hz mode (#844)

- Add decode-driven flow control and bounded GOP-aware queues to prevent stale H.264 frames from accumulating in TCP and WebSocket buffers.
- Isolate slow clients with dedicated writer goroutines, write deadlines, and safer connection lifecycle handling.
- Move the Direct H.264 socket into the worker, simplify the low-latency decode path, and improve reconnect behavior.
This commit is contained in:
watermeko
2026-07-31 19:08:07 +08:00
committed by GitHub
parent 8773605715
commit 21e834b4c8
7 changed files with 644 additions and 169 deletions

View File

@@ -0,0 +1,348 @@
package direct
import (
"encoding/binary"
"sync"
"time"
"github.com/gorilla/websocket"
log "github.com/sirupsen/logrus"
)
const (
frameAckMessage byte = 2
streamResyncMessage byte = 3
defaultQueueFrames = 8
defaultQueueBytes = 2 * 1024 * 1024
maxFlowWindow = 8
writeWait = 2 * time.Second
pingPeriod = 15 * time.Second
)
type outboundFrame struct {
key bool
timestamp int64
payload []byte
}
type frameQueue struct {
mutex sync.Mutex
wake chan struct{}
frames []*outboundFrame
queuedBytes int
inFlight []int64
maxFrames int
maxBytes int
window int
flowControlled bool
waitingForKeyframe bool
closed bool
}
func newFrameQueue(maxFrames int, maxBytes int) *frameQueue {
return &frameQueue{
wake: make(chan struct{}, 1),
maxFrames: maxFrames,
maxBytes: maxBytes,
waitingForKeyframe: true,
}
}
func (q *frameQueue) enableFlowControl(window int) {
if window < 1 {
window = 1
}
if window > maxFlowWindow {
window = maxFlowWindow
}
q.mutex.Lock()
q.clearFramesLocked()
q.inFlight = q.inFlight[:0]
q.window = window
q.flowControlled = true
q.waitingForKeyframe = true
q.mutex.Unlock()
}
func (q *frameQueue) canAdvanceStream() bool {
q.mutex.Lock()
defer q.mutex.Unlock()
if q.closed {
return false
}
if !q.flowControlled {
return true
}
if q.waitingForKeyframe {
return len(q.inFlight) < q.window
}
return len(q.frames)+len(q.inFlight) < q.window
}
func (q *frameQueue) captureState() (active bool, flowControlled bool, canAdvance bool) {
q.mutex.Lock()
defer q.mutex.Unlock()
if q.closed {
return false, q.flowControlled, false
}
if !q.flowControlled {
return true, false, true
}
if q.waitingForKeyframe {
return true, true, len(q.inFlight) < q.window
}
return true, true, len(q.frames)+len(q.inFlight) < q.window
}
func (q *frameQueue) offer(frame *outboundFrame) bool {
q.mutex.Lock()
defer q.mutex.Unlock()
if q.closed {
return false
}
if frame.key {
q.clearFramesLocked()
if q.flowControlled && len(q.inFlight) >= q.window {
q.waitingForKeyframe = true
return false
}
q.waitingForKeyframe = false
q.pushLocked(frame)
q.signalLocked()
return true
}
if q.waitingForKeyframe {
return false
}
if q.flowControlled {
if len(q.frames)+len(q.inFlight) >= q.window {
q.clearFramesLocked()
q.waitingForKeyframe = true
return false
}
} else if len(q.frames) >= q.maxFrames || q.queuedBytes+len(frame.payload) > q.maxBytes {
q.clearFramesLocked()
q.waitingForKeyframe = true
return false
}
q.pushLocked(frame)
q.signalLocked()
return true
}
func (q *frameQueue) popForWrite() *outboundFrame {
q.mutex.Lock()
defer q.mutex.Unlock()
if q.closed || len(q.frames) == 0 {
return nil
}
frame := q.frames[0]
q.frames[0] = nil
q.frames = q.frames[1:]
q.queuedBytes -= len(frame.payload)
if q.flowControlled {
q.inFlight = append(q.inFlight, frame.timestamp)
}
return frame
}
func (q *frameQueue) acknowledge(timestamp int64) {
q.mutex.Lock()
defer q.mutex.Unlock()
acknowledged := 0
for acknowledged < len(q.inFlight) && q.inFlight[acknowledged] <= timestamp {
acknowledged++
}
if acknowledged == 0 {
return
}
clear(q.inFlight[:acknowledged])
q.inFlight = q.inFlight[acknowledged:]
}
func (q *frameQueue) requestResync() {
q.mutex.Lock()
q.clearFramesLocked()
q.inFlight = q.inFlight[:0]
q.waitingForKeyframe = true
q.mutex.Unlock()
}
func (q *frameQueue) close() {
q.mutex.Lock()
q.closed = true
q.clearFramesLocked()
q.inFlight = q.inFlight[:0]
q.mutex.Unlock()
}
func (q *frameQueue) pushLocked(frame *outboundFrame) {
q.frames = append(q.frames, frame)
q.queuedBytes += len(frame.payload)
}
func (q *frameQueue) clearFramesLocked() {
clear(q.frames)
q.frames = q.frames[:0]
q.queuedBytes = 0
}
func (q *frameQueue) signalLocked() {
select {
case q.wake <- struct{}{}:
default:
}
}
type client struct {
conn *websocket.Conn
queue *frameQueue
done chan struct{}
writerDone chan struct{}
closeOnce sync.Once
}
func newClient(conn *websocket.Conn) *client {
return &client{
conn: conn,
queue: newFrameQueue(defaultQueueFrames, defaultQueueBytes),
done: make(chan struct{}),
writerDone: make(chan struct{}),
}
}
func (c *client) start() {
go c.writeLoop()
}
func (c *client) close() {
c.closeOnce.Do(func() {
close(c.done)
c.queue.close()
_ = c.conn.Close()
})
}
func (c *client) wait() {
<-c.writerDone
}
func (c *client) captureState() (active bool, flowControlled bool, canAdvance bool) {
return c.queue.captureState()
}
func hasCaptureDemand(clients []*client) bool {
for _, client := range clients {
active, flowControlled, canAdvance := client.captureState()
if !active {
continue
}
if !flowControlled || canAdvance {
return true
}
}
return false
}
func (c *client) offer(frame *outboundFrame) {
c.queue.offer(frame)
}
func (c *client) handleControl(messageType int, data []byte) {
if messageType != websocket.BinaryMessage || len(data) == 0 {
return
}
switch data[0] {
case frameAckMessage:
if len(data) == 9 {
timestamp := int64(binary.LittleEndian.Uint64(data[1:]))
c.queue.acknowledge(timestamp)
}
case streamResyncMessage:
if len(data) == 1 {
c.queue.requestResync()
}
}
}
func newOutboundFrame(isKeyFrame bool, timestamp int64, data []byte) *outboundFrame {
payload := make([]byte, 9+len(data))
if isKeyFrame {
payload[0] = 1
}
binary.LittleEndian.PutUint64(payload[1:9], uint64(timestamp))
copy(payload[9:], data)
return &outboundFrame{
key: isKeyFrame,
timestamp: timestamp,
payload: payload,
}
}
func (c *client) writeLoop() {
defer close(c.writerDone)
pingTicker := time.NewTicker(pingPeriod)
defer pingTicker.Stop()
for {
select {
case <-pingTicker.C:
deadline := time.Now().Add(writeWait)
if err := c.conn.WriteControl(websocket.PingMessage, nil, deadline); err != nil {
c.close()
return
}
default:
}
if frame := c.queue.popForWrite(); frame != nil {
if err := c.conn.SetWriteDeadline(time.Now().Add(writeWait)); err != nil {
c.close()
return
}
if err := c.conn.WriteMessage(websocket.BinaryMessage, frame.payload); err != nil {
log.Debugf("failed to write h264 frame to %s: %s", c.conn.RemoteAddr(), err)
c.close()
return
}
continue
}
select {
case <-c.done:
return
case <-c.queue.wake:
case <-pingTicker.C:
deadline := time.Now().Add(writeWait)
if err := c.conn.WriteControl(websocket.PingMessage, nil, deadline); err != nil {
c.close()
return
}
}
}
}

View File

@@ -3,6 +3,7 @@ package direct
import (
"NanoKVM-Server/service/stream"
"net/http"
"strconv"
"time"
"github.com/gin-gonic/gin"
@@ -10,6 +11,8 @@ import (
log "github.com/sirupsen/logrus"
)
const pongWait = 30 * time.Second
var (
streamer = newStreamer()
upgrader = websocket.Upgrader{
@@ -26,24 +29,36 @@ func Connect(c *gin.Context) {
log.Errorf("failed to upgrade to websocket: %s", err)
return
}
client := newClient(ws)
if flowWindow, err := strconv.Atoi(c.Query("flow")); err == nil && flowWindow > 0 {
client.queue.enableFlowControl(flowWindow)
}
defer func() {
_ = ws.Close()
streamer.removeClient(client)
client.close()
client.wait()
log.Debugf("h264 websocket disconnected: %s", ws.RemoteAddr())
}()
log.Debugf("h264 websocket connected: %s", ws.RemoteAddr())
_ = ws.SetReadDeadline(time.Time{})
ws.SetReadLimit(64)
_ = ws.SetReadDeadline(time.Now().Add(pongWait))
ws.SetPongHandler(func(string) error {
return ws.SetReadDeadline(time.Now().Add(pongWait))
})
streamer.addClient(ws)
defer streamer.removeClient(ws)
streamer.addClient(client)
unregisterMode := stream.RegisterH264Mode(stream.H264ModeDirect)
defer unregisterMode()
for {
if _, _, err := ws.NextReader(); err != nil {
messageType, data, err := ws.ReadMessage()
if err != nil {
log.Debugf("failed to read message (client disconnected): %s", err)
return
}
_ = ws.SetReadDeadline(time.Now().Add(pongWait))
client.handleControl(messageType, data)
}
}

View File

@@ -1,12 +0,0 @@
package direct
import (
"bytes"
"sync"
)
var BufferPool = sync.Pool{
New: func() interface{} {
return new(bytes.Buffer)
},
}

View File

@@ -4,62 +4,71 @@ import (
"NanoKVM-Server/common"
"NanoKVM-Server/service/stream"
"NanoKVM-Server/service/vm"
"bytes"
"encoding/binary"
"sync"
"sync/atomic"
"time"
"github.com/gorilla/websocket"
log "github.com/sirupsen/logrus"
)
type Streamer struct {
mutex sync.Mutex
clients map[*websocket.Conn]bool
clientSnapshot atomic.Pointer[[]*websocket.Conn]
running int32
clients map[*client]struct{}
clientSnapshot atomic.Pointer[[]*client]
running bool
viewerVersion uint64
}
func newStreamer() *Streamer {
s := &Streamer{
clients: make(map[*websocket.Conn]bool),
clients: make(map[*client]struct{}),
}
s.updateClientSnapshotLocked()
return s
}
func (s *Streamer) addClient(ws *websocket.Conn) {
func (s *Streamer) addClient(client *client) {
client.start()
s.mutex.Lock()
s.clients[ws] = true
s.clients[client] = struct{}{}
count := s.updateClientSnapshotLocked()
s.viewerVersion++
version := s.viewerVersion
start := !s.running
if start {
s.running = true
}
s.mutex.Unlock()
vm.UpdateHdmiViewerSnapshot("direct", count, version)
if atomic.CompareAndSwapInt32(&s.running, 0, 1) {
if start {
go s.run()
log.Debug("h264 stream started")
}
}
func (s *Streamer) removeClient(ws *websocket.Conn) {
func (s *Streamer) removeClient(client *client) {
s.mutex.Lock()
delete(s.clients, ws)
if _, exists := s.clients[client]; !exists {
s.mutex.Unlock()
return
}
delete(s.clients, client)
count := s.updateClientSnapshotLocked()
s.viewerVersion++
version := s.viewerVersion
s.mutex.Unlock()
client.close()
vm.UpdateHdmiViewerSnapshot("direct", count, version)
log.Debugf("h264 websocket disconnected, remaining clients: %d", count)
}
func (s *Streamer) updateClientSnapshotLocked() int {
clients := make([]*websocket.Conn, 0, len(s.clients))
clients := make([]*client, 0, len(s.clients))
for client := range s.clients {
clients = append(clients, client)
}
@@ -68,7 +77,7 @@ func (s *Streamer) updateClientSnapshotLocked() int {
return len(clients)
}
func (s *Streamer) getClients() []*websocket.Conn {
func (s *Streamer) getClients() []*client {
clients := s.clientSnapshot.Load()
if clients == nil {
return nil
@@ -78,8 +87,6 @@ func (s *Streamer) getClients() []*websocket.Conn {
}
func (s *Streamer) run() {
defer atomic.StoreInt32(&s.running, 0)
screen := common.GetScreen()
common.CheckScreen()
fps := screen.FPS
@@ -93,24 +100,10 @@ func (s *Streamer) run() {
for range ticker.C {
clients := s.getClients()
if len(clients) == 0 {
log.Debug("h264 stream stopped due to no clients")
return
}
data, result := vision.ReadH264(screen.Width, screen.Height, screen.BitRate)
stream.UpdateCaptureStatus(stream.CaptureModeDirect, result)
if result < 0 || len(data) == 0 {
continue
}
isKeyFrame := byte(0)
if result == 3 {
isKeyFrame = byte(1)
}
timestamp := time.Since(startTime).Microseconds()
if err := s.send(clients, isKeyFrame, timestamp, data); err != nil {
if s.stopIfIdle() {
log.Debug("h264 stream stopped due to no clients")
return
}
continue
}
@@ -119,42 +112,34 @@ func (s *Streamer) run() {
ticker.Reset(time.Second / time.Duration(fps))
}
if !hasCaptureDemand(clients) {
continue
}
data, result := vision.ReadH264(screen.Width, screen.Height, screen.BitRate)
stream.UpdateCaptureStatus(stream.CaptureModeDirect, result)
if result < 0 || len(data) == 0 {
continue
}
timestamp := time.Since(startTime).Microseconds()
frame := newOutboundFrame(result == 3, timestamp, data)
for _, client := range clients {
client.offer(frame)
}
stream.GetFrameRateCounter().Update()
}
}
func (s *Streamer) send(clients []*websocket.Conn, isKeyFrame byte, timestamp int64, data []byte) error {
buf := BufferPool.Get().(*bytes.Buffer)
defer BufferPool.Put(buf)
func (s *Streamer) stopIfIdle() bool {
s.mutex.Lock()
defer s.mutex.Unlock()
buf.Reset()
buf.Grow(1 + 8 + len(data))
if err := buf.WriteByte(isKeyFrame); err != nil {
log.Errorf("failed to write keyframe flag: %s", err)
return err
if len(s.clients) > 0 {
return false
}
var tsBytes [8]byte
binary.LittleEndian.PutUint64(tsBytes[:], uint64(timestamp))
if _, err := buf.Write(tsBytes[:]); err != nil {
log.Errorf("failed to write timestamp: %s", err)
return err
}
if _, err := buf.Write(data); err != nil {
log.Errorf("failed to write h264 data: %s", err)
return err
}
for _, client := range clients {
if err := client.WriteMessage(websocket.BinaryMessage, buf.Bytes()); err != nil {
log.Errorf("failed to write message to client %s: %s.", client.RemoteAddr(), err)
s.removeClient(client)
_ = client.Close()
}
}
return nil
s.running = false
return true
}