Fix packet forwarding between vz and socket_vmnet

We used external package (tcpproxy) for proxying between unix stream and
datagram sockets. This package cannot handle ENOBUFS error, expected
condition on BSD based systems, and worse, it hides errors and stop
forwarding packets silently when write to vz socket fails with
ENOBUFS[1].

Fix the issues by replacing tcpproxy with a simpler and more direct
implementation that will be easier to maintain.

Fixes:

- Fix error handling if write to vz datagram socket fail with ENOBUFS.
  We retry the write until it succeeds with a very short sleep between
  retries. Similar solution is used in gvisor-tap-vsock[2].

- Fix error handling if we could not read packet header or body from
  socket_vmnet stream socket. Previously we logged an error and continue
  to send corrupted packet to vz from the point of the failure.

- Fix error handling if writing a packet to socket_vmnet stream socket
  returned after writing partial packet. Now we handle short writes and
  write the complete packet. Previously would break the protocol and
  continue to send corrupted packet from the point of the failure.

- Log error if forwarding packets from vz to socket_vmnet or from
  socket_vmnet to vz failed.

Simplification:

- Use binary.Read() and binary.Write() to read and write qemu packet
  header.

Visibility:

- Make QEMUPacketConn private since it is an implementation detail of vz
  when using socket_vmnet.

[1] lima-vm/socket_vmnet#39
[2] containers/gvisor-tap-vsock#370

Signed-off-by: Nir Soffer <[email protected]>

go.mod

@@ -47,7 +47,7 @@ require (
 	google.golang.org/protobuf v1.34.2
 	gopkg.in/op/go-logging.v1 v1.0.0-20160211212156-b2cb9fa56473
 	gotest.tools/v3 v3.5.1
-	inet.af/tcpproxy v0.0.0-20221017015627-91f861402626 // replaced to github.com/inetaf/tcpproxy (see the bottom of this go.mod file)
+	inet.af/tcpproxy v0.0.0-20221017015627-91f861402626 // indirect; replaced to github.com/inetaf/tcpproxy (see the bottom of this go.mod file)
 	k8s.io/api v0.31.1
 	k8s.io/apimachinery v0.31.1
 	k8s.io/client-go v0.31.1

pkg/vz/network_darwin.go

@@ -3,17 +3,18 @@
 package vz
 
 import (
-	"context"
 	"encoding/binary"
+	"errors"
 	"io"
 	"net"
 	"os"
+	"sync"
+	"syscall"
 	"time"
 
 	"github.com/balajiv113/fd"
 
 	"github.com/sirupsen/logrus"
-	"inet.af/tcpproxy" // replaced to github.com/inetaf/tcpproxy in go.mod
 )
 
 func PassFDToUnix(unixSock string) (*os.File, error) {
@@ -40,7 +41,7 @@ func DialQemu(unixSock string) (*os.File, error) {
 	if err != nil {
 		return nil, err
 	}
-	qemuConn := &QEMUPacketConn{unixConn: unixConn}
+	qemuConn := &qemuPacketConn{Conn: unixConn}
 
 	server, client, err := createSockPair()
 	if err != nil {
@@ -50,77 +51,117 @@ func DialQemu(unixSock string) (*os.File, error) {
 	if err != nil {
 		return nil, err
 	}
+	vzConn := &packetConn{Conn: dgramConn}
 
-	remote := tcpproxy.DialProxy{
-		DialContext: func(context.Context, string, string) (net.Conn, error) {
-			return dgramConn, nil
-		},
-	}
-	go remote.HandleConn(qemuConn)
+	go forwardPackets(qemuConn, vzConn)
 
 	return client, nil
 }
 
-// QEMUPacketConn converts raw network packet to a QEMU supported network packet.
-type QEMUPacketConn struct {
-	unixConn net.Conn
+func forwardPackets(qemuConn *qemuPacketConn, vzConn *packetConn) {
+	defer qemuConn.Close()
+	defer vzConn.Close()
+
+	var wg sync.WaitGroup
+	wg.Add(2)
+
+	go func() {
+		defer wg.Done()
+		if _, err := io.Copy(qemuConn, vzConn); err != nil {
+			logrus.Errorf("Failed to forward packets from VZ to VMNET: %s", err)
+		}
+	}()
+
+	go func() {
+		defer wg.Done()
+		if _, err := io.Copy(vzConn, qemuConn); err != nil {
+			logrus.Errorf("Failed to forward packets from VMNET to VZ: %s", err)
+		}
+	}()
+
+	wg.Wait()
 }
 
-var _ net.Conn = (*QEMUPacketConn)(nil)
+// qemuPacketConn converts raw network packet to a QEMU supported network packet.
+type qemuPacketConn struct {
+	net.Conn
+}
 
-// Read gets rid of the QEMU header packet and returns the raw packet as response.
-func (v *QEMUPacketConn) Read(b []byte) (n int, err error) {
-	header := make([]byte, 4)
-	_, err = io.ReadFull(v.unixConn, header)
-	if err != nil {
-		logrus.Errorln("Failed to read header", err)
+// Read reads a QEMU packet and returns the contained raw packet.  Returns (len,
+// nil) if a packet was read, and (0, err) on error. Errors means the prorocol
+// is broken and the socket must be closed.
+func (c *qemuPacketConn) Read(b []byte) (n int, err error) {
+	var size uint32
+	if err := binary.Read(c.Conn, binary.BigEndian, &size); err != nil {
+		// Likely connection closed by peer.
+		return 0, err
 	}
 
-	size := binary.BigEndian.Uint32(header)
-	reader := io.LimitReader(v.unixConn, int64(size))
+	reader := io.LimitReader(c.Conn, int64(size))
 	_, err = reader.Read(b)
 	if err != nil {
-		logrus.Errorln("Failed to read packet", err)
+		// Likely connection closed by peer.
+		return 0, err
 	}
 	return int(size), nil
 }
 
-// Write puts QEMU header packet first and then writes the raw packet.
-func (v *QEMUPacketConn) Write(b []byte) (n int, err error) {
-	header := make([]byte, 4)
-	binary.BigEndian.PutUint32(header, uint32(len(b)))
-	_, err = v.unixConn.Write(header)
-	if err != nil {
-		logrus.Errorln("Failed to write header", err)
+// Write writes a QEMU packet containing the raw packet. Returns (len(b), nil)
+// if a packet was written, and (0, err) if a packet was not fully written.
+// Errors means the prorocol is broken and the socket must be closed.
+func (c *qemuPacketConn) Write(b []byte) (int, error) {
+	size := len(b)
+	header := uint32(size)
+	if err := binary.Write(c.Conn, binary.BigEndian, header); err != nil {
+		return 0, err
 	}
 
-	write, err := v.unixConn.Write(b)
-	if err != nil {
-		logrus.Errorln("Failed to write packet", err)
+	start := 0
+	for start < size {
+		nw, err := c.Conn.Write(b[start:])
+		if err != nil {
+			return 0, err
+		}
+		start += nw
 	}
-	return write, nil
+	return size, nil
 }
 
-func (v *QEMUPacketConn) Close() error {
-	return v.unixConn.Close()
-}
-
-func (v *QEMUPacketConn) LocalAddr() net.Addr {
-	return v.unixConn.LocalAddr()
-}
-
-func (v *QEMUPacketConn) RemoteAddr() net.Addr {
-	return v.unixConn.RemoteAddr()
-}
-
-func (v *QEMUPacketConn) SetDeadline(t time.Time) error {
-	return v.unixConn.SetDeadline(t)
-}
+// Testing show that retries are very rare (e.g 24 of 62,499,008 packets) and
+// requires 1 or 2 retries to complete the write. A 100 microseconds sleep loop
+// consumes about 4% CPU on M1 Pro.
+const writeRetryDelay = 100 * time.Microsecond
 
-func (v *QEMUPacketConn) SetReadDeadline(t time.Time) error {
-	return v.unixConn.SetReadDeadline(t)
+// packetConn handles ENOBUFS errors when writing to unixgram socket.
+type packetConn struct {
+	net.Conn
 }
 
-func (v *QEMUPacketConn) SetWriteDeadline(t time.Time) error {
-	return v.unixConn.SetWriteDeadline(t)
+// Write write a packet retrying on ENOBUFS errors.
+func (c *packetConn) Write(b []byte) (int, error) {
+	var retries uint64
+	for {
+		n, err := c.Conn.Write(b)
+		if err != nil && errors.Is(err, syscall.ENOBUFS) {
+			// This is an expected condition on BSD based system. The kernel
+			// does not support blocking until buffer space is available.
+			// The only way to recover is to retry the call until it
+			// succeeds, or drop the packet.
+			// Handled in a similar way in gvisor-tap-vsock:
+			// https://github.com/containers/gvisor-tap-vsock/issues/367
+			time.Sleep(writeRetryDelay)
+			retries++
+			continue
+		}
+		if err != nil {
+			return 0, err
+		}
+		if n < len(b) {
+			return n, errors.New("incomplete write to unixgram socket")
+		}
+		if retries > 0 {
+			logrus.Debugf("Write completed after %d retries", retries)
+		}
+		return n, nil
+	}
 }