From a2234b6a2ae2eba7ae95d57c9d90a7facc553844 Mon Sep 17 00:00:00 2001
From: zarazaex69 <zarazaex@tuta.io>
Date: Sun, 3 May 2026 16:10:32 +0300
Subject: [PATCH] perf: optimize write polling and kcp interval for lower
 latency

---
 internal/muxconn/conn.go             | 17 +++++++++++++++--
 internal/transport/vp8channel/kcp.go | 13 ++++++++-----
 2 files changed, 23 insertions(+), 7 deletions(-)

diff --git a/internal/muxconn/conn.go b/internal/muxconn/conn.go
index 3f651a5..c5d6088 100644
--- a/internal/muxconn/conn.go
+++ b/internal/muxconn/conn.go
@@ -19,6 +19,7 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"runtime"
 	"sync"
 	"time"
 
@@ -81,14 +82,26 @@ func (c *Conn) Read(p []byte) (int, error) {
 // Write encrypts p and ships it to the link as a single message. Blocks while
 // the link signals back-pressure.
 func (c *Conn) Write(p []byte) (int, error) {
-	for {
+	// Spin briefly first - on a healthy link CanSend usually clears within
+	// well under a millisecond, so a 10ms sleep adds visible per-frame
+	// latency to interactive request/response traffic. Fall back to a
+	// modest sleep only if the link is truly congested.
+	const (
+		fastSpinAttempts = 200
+		slowPollDelay    = 2 * time.Millisecond
+	)
+	for attempt := 0; ; attempt++ {
 		if c.isClosed() {
 			return 0, ErrClosed
 		}
 		if c.ln.CanSend() {
 			break
 		}
-		time.Sleep(10 * time.Millisecond)
+		if attempt < fastSpinAttempts {
+			runtime.Gosched()
+			continue
+		}
+		time.Sleep(slowPollDelay)
 	}
 
 	enc, err := c.cipher.Encrypt(p)
diff --git a/internal/transport/vp8channel/kcp.go b/internal/transport/vp8channel/kcp.go
index f3988ef..9140586 100644
--- a/internal/transport/vp8channel/kcp.go
+++ b/internal/transport/vp8channel/kcp.go
@@ -65,11 +65,14 @@ func startKCP(out chan<- []byte, onData func([]byte), epochHdr [epochHdrLen]byte
 		return nil, fmt.Errorf("kcp new conn: %w", err)
 	}
 
-	// Aggressive ARQ tuning: nodelay=1, interval=10ms, fast resend=2, no
-	// congestion control. This is the standard "turbo" preset used by KCP
-	// tunnels on lossy networks (shadowsocks, kcptun) and is the whole point
-	// of choosing KCP over SCTP.
-	sess.SetNoDelay(1, 10, 2, 1)
+	// Aggressive ARQ tuning: nodelay=1, interval=5ms, fast resend=2, no
+	// congestion control. The 5ms tick (vs the kcptun-default 10ms) halves
+	// the worst-case scheduling latency in each direction, which matters a
+	// lot for interactive workloads (SOCKS5 + HTTP request needs ~3 RTTs
+	// before the first byte of the response shows up). Below 5ms the
+	// CPU cost of the KCP update loop starts climbing without much
+	// additional latency benefit.
+	sess.SetNoDelay(1, 5, 2, 1)
 	sess.SetWindowSize(kcpSndWnd, kcpRcvWnd)
 	sess.SetMtu(kcpMTU)
 	sess.SetACKNoDelay(true)