nebula/udp/udp_linux.go

//go:build !android && !e2e_testing
// +build !android,!e2e_testing

package udp

import (
	"encoding/binary"
	"errors"
	"fmt"
	"log/slog"
	"net"
	"net/netip"
	"sync/atomic"
	"syscall"
	"unsafe"

	"github.com/rcrowley/go-metrics"
	"github.com/slackhq/nebula/config"
	"golang.org/x/sys/unix"
)

type StdConn struct {
	sysFd  int
	closed atomic.Bool
	isV4   bool
	l      *slog.Logger
	batch  int
}

func NewListener(l *slog.Logger, ip netip.Addr, port int, multi bool, batch int) (Conn, error) {
	af := unix.AF_INET6
	if ip.Is4() {
		af = unix.AF_INET
	}
	syscall.ForkLock.RLock()
	fd, err := unix.Socket(af, unix.SOCK_DGRAM, unix.IPPROTO_UDP)
	if err == nil {
		unix.CloseOnExec(fd)
	}
	syscall.ForkLock.RUnlock()
	if err != nil {
		return nil, fmt.Errorf("unable to open socket: %w", err)
	}

	if multi {
		if err = unix.SetsockoptInt(fd, unix.SOL_SOCKET, unix.SO_REUSEPORT, 1); err != nil {
			_ = unix.Close(fd)
			return nil, fmt.Errorf("unable to set SO_REUSEPORT: %w", err)
		}
	}

	var sa unix.Sockaddr
	if ip.Is4() {
		sa4 := &unix.SockaddrInet4{Port: port}
		sa4.Addr = ip.As4()
		sa = sa4
	} else {
		sa6 := &unix.SockaddrInet6{Port: port}
		sa6.Addr = ip.As16()
		sa = sa6
	}
	if err = unix.Bind(fd, sa); err != nil {
		_ = unix.Close(fd)
		return nil, fmt.Errorf("unable to bind to socket: %w", err)
	}

	return &StdConn{sysFd: fd, isV4: ip.Is4(), l: l, batch: batch}, nil
}

func (u *StdConn) SupportsMultipleReaders() bool {
	return true
}

func (u *StdConn) Rebind() error {
	return nil
}

func (u *StdConn) SetRecvBuffer(n int) error {
	return unix.SetsockoptInt(u.sysFd, unix.SOL_SOCKET, unix.SO_RCVBUFFORCE, n)
}

func (u *StdConn) SetSendBuffer(n int) error {
	return unix.SetsockoptInt(u.sysFd, unix.SOL_SOCKET, unix.SO_SNDBUFFORCE, n)
}

func (u *StdConn) SetSoMark(mark int) error {
	return unix.SetsockoptInt(u.sysFd, unix.SOL_SOCKET, unix.SO_MARK, mark)
}

func (u *StdConn) GetRecvBuffer() (int, error) {
	return unix.GetsockoptInt(u.sysFd, unix.SOL_SOCKET, unix.SO_RCVBUF)
}

func (u *StdConn) GetSendBuffer() (int, error) {
	return unix.GetsockoptInt(u.sysFd, unix.SOL_SOCKET, unix.SO_SNDBUF)
}

func (u *StdConn) GetSoMark() (int, error) {
	return unix.GetsockoptInt(u.sysFd, unix.SOL_SOCKET, unix.SO_MARK)
}

func (u *StdConn) LocalAddr() (netip.AddrPort, error) {
	sa, err := unix.Getsockname(u.sysFd)
	if err != nil {
		return netip.AddrPort{}, err
	}
	switch sa := sa.(type) {
	case *unix.SockaddrInet4:
		return netip.AddrPortFrom(netip.AddrFrom4(sa.Addr), uint16(sa.Port)), nil
	case *unix.SockaddrInet6:
		return netip.AddrPortFrom(netip.AddrFrom16(sa.Addr), uint16(sa.Port)), nil
	default:
		return netip.AddrPort{}, fmt.Errorf("unsupported sock type: %T", sa)
	}
}

// recvmmsg does one blocking recvmmsg (MSG_WAITFORONE), reading up to len(msgs) datagrams
func (u *StdConn) recvmmsg(msgs []rawMessage) (int, error) {
	r, _, errno := unix.Syscall6(
		unix.SYS_RECVMMSG,
		uintptr(u.sysFd),
		uintptr(unsafe.Pointer(&msgs[0])),
		uintptr(len(msgs)),
		unix.MSG_WAITFORONE,
		0,
		0,
	)
	if errno != 0 {
		if u.closed.Load() {
			return 0, net.ErrClosed
		}
		return 0, &net.OpError{Op: "recvmmsg", Err: errno}
	}
	n := int(r)
	if (n == 0 || msgs[0].Len == 0) && u.closed.Load() {
		return 0, net.ErrClosed
	}
	return n, nil
}

// recvmsg does one blocking recvmsg into msgs[0]
func (u *StdConn) recvmsg(msgs []rawMessage) (int, error) {
	r, _, errno := unix.Syscall6(
		unix.SYS_RECVMSG,
		uintptr(u.sysFd),
		uintptr(unsafe.Pointer(&msgs[0].Hdr)),
		0,
		0,
		0,
		0,
	)
	if errno != 0 {
		if u.closed.Load() {
			return 0, net.ErrClosed
		}
		return 0, &net.OpError{Op: "recvmsg", Err: errno}
	}
	if r == 0 && u.closed.Load() {
		return 0, net.ErrClosed
	}
	msgs[0].Len = uint32(r)
	return 1, nil
}

func (u *StdConn) ListenOut(r EncReader) error {
	var ip netip.Addr
	msgs, buffers, names := u.PrepareRawMessages(u.batch)
	read := u.recvmmsg
	if u.batch == 1 {
		read = u.recvmsg
	}

	for {
		n, err := read(msgs)
		if err != nil {
			if errors.Is(err, unix.EINTR) {
				continue // interrupted by a signal, retry the read
			}
			// net.ErrClosed after Close() is teardown, absorbed by the caller's
			// closed flag like the other platforms; anything else is a real error.
			return err
		}

		for i := 0; i < n; i++ {
			// Its ok to skip the ok check here, the slicing is the only error that can occur and it will panic
			if u.isV4 {
				ip, _ = netip.AddrFromSlice(names[i][4:8])
			} else {
				ip, _ = netip.AddrFromSlice(names[i][8:24])
			}
			r(netip.AddrPortFrom(ip.Unmap(), binary.BigEndian.Uint16(names[i][2:4])), buffers[i][:msgs[i].Len])
		}
	}
}

func (u *StdConn) WriteTo(b []byte, ip netip.AddrPort) error {
	if u.isV4 {
		return u.writeTo4(b, ip)
	}
	return u.writeTo6(b, ip)
}

func (u *StdConn) writeTo6(b []byte, ip netip.AddrPort) error {
	var rsa unix.RawSockaddrInet6
	rsa.Family = unix.AF_INET6
	rsa.Addr = ip.Addr().As16()
	binary.BigEndian.PutUint16((*[2]byte)(unsafe.Pointer(&rsa.Port))[:], ip.Port())

	for {
		_, _, err := unix.Syscall6(
			unix.SYS_SENDTO,
			uintptr(u.sysFd),
			uintptr(unsafe.Pointer(&b[0])),
			uintptr(len(b)),
			uintptr(0),
			uintptr(unsafe.Pointer(&rsa)),
			uintptr(unix.SizeofSockaddrInet6),
		)
		if err != 0 {
			return &net.OpError{Op: "sendto", Err: err}
		}
		return nil
	}
}

func (u *StdConn) writeTo4(b []byte, ip netip.AddrPort) error {
	if !ip.Addr().Is4() {
		return ErrInvalidIPv6RemoteForSocket
	}

	var rsa unix.RawSockaddrInet4
	rsa.Family = unix.AF_INET
	rsa.Addr = ip.Addr().As4()
	binary.BigEndian.PutUint16((*[2]byte)(unsafe.Pointer(&rsa.Port))[:], ip.Port())

	for {
		_, _, err := unix.Syscall6(
			unix.SYS_SENDTO,
			uintptr(u.sysFd),
			uintptr(unsafe.Pointer(&b[0])),
			uintptr(len(b)),
			uintptr(0),
			uintptr(unsafe.Pointer(&rsa)),
			uintptr(unix.SizeofSockaddrInet4),
		)
		if err != 0 {
			return &net.OpError{Op: "sendto", Err: err}
		}
		return nil
	}
}

func (u *StdConn) ReloadConfig(c *config.C) {
	b := c.GetInt("listen.read_buffer", 0)
	if b > 0 {
		if err := u.SetRecvBuffer(b); err == nil {
			if s, err := u.GetRecvBuffer(); err == nil {
				u.l.Info("listen.read_buffer was set", "size", s)
			} else {
				u.l.Warn("Failed to get listen.read_buffer", "error", err)
			}
		} else {
			u.l.Error("Failed to set listen.read_buffer", "error", err)
		}
	}

	b = c.GetInt("listen.write_buffer", 0)
	if b > 0 {
		if err := u.SetSendBuffer(b); err == nil {
			if s, err := u.GetSendBuffer(); err == nil {
				u.l.Info("listen.write_buffer was set", "size", s)
			} else {
				u.l.Warn("Failed to get listen.write_buffer", "error", err)
			}
		} else {
			u.l.Error("Failed to set listen.write_buffer", "error", err)
		}
	}

	b = c.GetInt("listen.so_mark", 0)
	s, err := u.GetSoMark()
	if b > 0 || (err == nil && s != 0) {
		if err := u.SetSoMark(b); err == nil {
			if s, err := u.GetSoMark(); err == nil {
				u.l.Info("listen.so_mark was set", "mark", s)
			} else {
				u.l.Warn("Failed to get listen.so_mark", "error", err)
			}
		} else {
			u.l.Error("Failed to set listen.so_mark", "error", err)
		}
	}
}

func (u *StdConn) getMemInfo(meminfo *[unix.SK_MEMINFO_VARS]uint32) error {
	var vallen uint32 = 4 * unix.SK_MEMINFO_VARS
	_, _, err := unix.Syscall6(unix.SYS_GETSOCKOPT, uintptr(u.sysFd), uintptr(unix.SOL_SOCKET), uintptr(unix.SO_MEMINFO), uintptr(unsafe.Pointer(meminfo)), uintptr(unsafe.Pointer(&vallen)), 0)
	if err != 0 {
		return err
	}
	return nil
}

func (u *StdConn) Close() error {
	u.closed.Store(true)
	// Wake the reader parked in recvmmsg/recvmsg. shutdown(2) on an unconnected socket
	// returns ENOTCONN but still wakes it, so ignore the error.
	// The reader then sees closed and stops touching the fd, making the Close below safe.
	_ = unix.Shutdown(u.sysFd, unix.SHUT_RDWR)
	return unix.Close(u.sysFd)
}

func NewUDPStatsEmitter(udpConns []Conn) func() {
	// Check if our kernel supports SO_MEMINFO before registering the gauges
	var udpGauges [][unix.SK_MEMINFO_VARS]metrics.Gauge
	var meminfo [unix.SK_MEMINFO_VARS]uint32
	if err := udpConns[0].(*StdConn).getMemInfo(&meminfo); err == nil {
		udpGauges = make([][unix.SK_MEMINFO_VARS]metrics.Gauge, len(udpConns))
		for i := range udpConns {
			udpGauges[i] = [unix.SK_MEMINFO_VARS]metrics.Gauge{
				metrics.GetOrRegisterGauge(fmt.Sprintf("udp.%d.rmem_alloc", i), nil),
				metrics.GetOrRegisterGauge(fmt.Sprintf("udp.%d.rcvbuf", i), nil),
				metrics.GetOrRegisterGauge(fmt.Sprintf("udp.%d.wmem_alloc", i), nil),
				metrics.GetOrRegisterGauge(fmt.Sprintf("udp.%d.sndbuf", i), nil),
				metrics.GetOrRegisterGauge(fmt.Sprintf("udp.%d.fwd_alloc", i), nil),
				metrics.GetOrRegisterGauge(fmt.Sprintf("udp.%d.wmem_queued", i), nil),
				metrics.GetOrRegisterGauge(fmt.Sprintf("udp.%d.optmem", i), nil),
				metrics.GetOrRegisterGauge(fmt.Sprintf("udp.%d.backlog", i), nil),
				metrics.GetOrRegisterGauge(fmt.Sprintf("udp.%d.drops", i), nil),
			}
		}
	}

	return func() {
		for i, gauges := range udpGauges {
			if err := udpConns[i].(*StdConn).getMemInfo(&meminfo); err == nil {
				for j := 0; j < unix.SK_MEMINFO_VARS; j++ {
					gauges[j].Update(int64(meminfo[j]))
				}
			}
		}
	}
}