ipn/ipnlocal: make Serve idle connection limit configurable (#21179)

* ipn/ipnlocal: make Serve idle connection limit configurable

Allow Kubernetes operator proxy pods to override the per-host idle connection limit through an environment knob.\n\nUpdates tailscale/tailscale#20875

Signed-off-by: maxiscoding28 <max.winslow@icloud.com>

* ipn/ipnlocal: document Serve idle connection default

Signed-off-by: maxiscoding28 <max.winslow@icloud.com>

---------

Signed-off-by: maxiscoding28 <max.winslow@icloud.com>

ipn/ipnlocal/serve.go

@@ -35,6 +35,7 @@
 
 	"github.com/pires/go-proxyproto"
 	"go4.org/mem"
+	"tailscale.com/envknob"
 	"tailscale.com/ipn"
 	"tailscale.com/net/netmon"
 	"tailscale.com/net/netutil"
@@ -1001,6 +1002,8 @@ func (rp *reverseProxy) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 // to the backend. The Transport gets created lazily, at most once.
 func (rp *reverseProxy) getTransport() *http.Transport {
 	return rp.httpTransport.Get(func() *http.Transport {
+		// Zero preserves http.Transport's default MaxIdleConnsPerHost value.
+		maxIdleConnsPerHost, _ := envknob.LookupInt("TS_DEBUG_SERVE_MAX_IDLE_CONNS_PER_HOST")
 		dial := rp.lb.dialer.SystemDial
 		if rp.socketPath != "" {
 			dial = func(ctx context.Context, _, _ string) (net.Conn, error) {
@@ -1017,6 +1020,7 @@ func (rp *reverseProxy) getTransport() *http.Transport {
 			// Values for the following parameters have been copied from http.DefaultTransport.
 			ForceAttemptHTTP2:     true,
 			MaxIdleConns:          100,
+			MaxIdleConnsPerHost:   maxIdleConnsPerHost,
 			IdleConnTimeout:       90 * time.Second,
 			TLSHandshakeTimeout:   10 * time.Second,
 			ExpectContinueTimeout: 1 * time.Second,

ipn/ipnlocal/serve_test.go

@@ -29,9 +29,12 @@
 	"time"
 
 	"tailscale.com/control/controlclient"
+	"tailscale.com/envknob"
 	"tailscale.com/health"
 	"tailscale.com/ipn"
 	"tailscale.com/ipn/store/mem"
+	"tailscale.com/net/netmon"
+	"tailscale.com/net/tsdial"
 	"tailscale.com/tailcfg"
 	"tailscale.com/tailcfg/nodecap"
 	"tailscale.com/tailcfg/peercap"
@@ -1090,6 +1093,25 @@ type test struct {
 	})
 }
 
+func TestServeMaxIdleConnsPerHost(t *testing.T) {
+	for _, tt := range []struct {
+		name string
+		env  string
+		want int
+	}{
+		{name: "default", want: 0},
+		{name: "configured", env: "100", want: 100},
+	} {
+		t.Run(tt.name, func(t *testing.T) {
+			envknob.SetenvForTest(t, "TS_DEBUG_SERVE_MAX_IDLE_CONNS_PER_HOST", tt.env)
+			rp := &reverseProxy{lb: &LocalBackend{dialer: tsdial.NewDialer(netmon.NewStatic())}}
+			if got := rp.getTransport().MaxIdleConnsPerHost; got != tt.want {
+				t.Errorf("MaxIdleConnsPerHost = %d, want %d", got, tt.want)
+			}
+		})
+	}
+}
+
 func mustCreateURL(t *testing.T, u string) url.URL {
 	t.Helper()
 	uParsed, err := url.Parse(u)