proxier.go

/*
Copyright 2014 Google Inc. All rights reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
*/

package proxy

import (
	"fmt"
	"io"
	"net"
	"strconv"
	"strings"
	"sync"
	"time"

	"github.com/GoogleCloudPlatform/kubernetes/pkg/api"
	"github.com/GoogleCloudPlatform/kubernetes/pkg/util"
	"github.com/GoogleCloudPlatform/kubernetes/pkg/util/errors"
	"github.com/GoogleCloudPlatform/kubernetes/pkg/util/iptables"
	"github.com/golang/glog"
)

type serviceInfo struct {
	portalIP   net.IP
	portalPort int
	protocol   api.Protocol
	proxyPort  int
	socket     proxySocket
	timeout    time.Duration
	// TODO: make this an net.IP address
	publicIP            []string
	sessionAffinityType api.AffinityType
	stickyMaxAgeMinutes int
}

// How long we wait for a connection to a backend in seconds
var endpointDialTimeout = []time.Duration{1, 2, 4, 8}

// Abstraction over TCP/UDP sockets which are proxied.
type proxySocket interface {
	// Addr gets the net.Addr for a proxySocket.
	Addr() net.Addr
	// Close stops the proxySocket from accepting incoming connections.  Each implementation should comment
	// on the impact of calling Close while sessions are active.
	Close() error
	// ProxyLoop proxies incoming connections for the specified service to the service endpoints.
	ProxyLoop(service string, proxier *Proxier)
}

// tcpProxySocket implements proxySocket.  Close() is implemented by net.Listener.  When Close() is called,
// no new connections are allowed but existing connections are left untouched.
type tcpProxySocket struct {
	net.Listener
}

func tryConnect(service string, srcAddr net.Addr, protocol string, proxier *Proxier) (out net.Conn, err error) {
	for _, retryTimeout := range endpointDialTimeout {
		endpoint, err := proxier.loadBalancer.NextEndpoint(service, srcAddr)
		if err != nil {
			glog.Errorf("Couldn't find an endpoint for %s: %v", service, err)
			return nil, err
		}
		glog.V(3).Infof("Mapped service %q to endpoint %s", service, endpoint)
		// TODO: This could spin up a new goroutine to make the outbound connection,
		// and keep accepting inbound traffic.
		outConn, err := net.DialTimeout(protocol, endpoint, retryTimeout*time.Second)
		if err != nil {
			glog.Errorf("Dial failed: %v", err)
			continue
		}
		return outConn, nil
	}
	return nil, fmt.Errorf("failed to connect to an endpoint.")
}

func (tcp *tcpProxySocket) ProxyLoop(service string, proxier *Proxier) {
	for {
		_, exists := proxier.getServiceInfo(service)
		if !exists {
			break
		}

		// Block until a connection is made.
		inConn, err := tcp.Accept()
		if err != nil {
			_, exists := proxier.getServiceInfo(service)
			if !exists {
				// Then the service port was just closed so the accept failure is to be expected.
				return
			}
			glog.Errorf("Accept failed: %v", err)
			continue
		}
		glog.V(2).Infof("Accepted TCP connection from %v to %v", inConn.RemoteAddr(), inConn.LocalAddr())
		outConn, err := tryConnect(service, inConn.(*net.TCPConn).RemoteAddr(), "tcp", proxier)
		if err != nil {
			glog.Errorf("Failed to connect to balancer: %v", err)
			inConn.Close()
			continue
		}
		// Spin up an async copy loop.
		go proxyTCP(inConn.(*net.TCPConn), outConn.(*net.TCPConn))
	}
}

// proxyTCP proxies data bi-directionally between in and out.
func proxyTCP(in, out *net.TCPConn) {
	var wg sync.WaitGroup
	wg.Add(2)
	glog.V(4).Infof("Creating proxy between %v <-> %v <-> %v <-> %v",
		in.RemoteAddr(), in.LocalAddr(), out.LocalAddr(), out.RemoteAddr())
	go copyBytes("from backend", in, out, &wg)
	go copyBytes("to backend", out, in, &wg)
	wg.Wait()
	in.Close()
	out.Close()
}

func copyBytes(direction string, dest, src *net.TCPConn, wg *sync.WaitGroup) {
	defer wg.Done()
	glog.V(4).Infof("Copying %s: %s -> %s", direction, src.RemoteAddr(), dest.RemoteAddr())
	n, err := io.Copy(dest, src)
	if err != nil {
		glog.Errorf("I/O error: %v", err)
	}
	glog.V(4).Infof("Copied %d bytes %s: %s -> %s", n, direction, src.RemoteAddr(), dest.RemoteAddr())
	dest.CloseWrite()
	src.CloseRead()
}

// udpProxySocket implements proxySocket.  Close() is implemented by net.UDPConn.  When Close() is called,
// no new connections are allowed and existing connections are broken.
// TODO: We could lame-duck this ourselves, if it becomes important.
type udpProxySocket struct {
	*net.UDPConn
}

func (udp *udpProxySocket) Addr() net.Addr {
	return udp.LocalAddr()
}

// Holds all the known UDP clients that have not timed out.
type clientCache struct {
	mu      sync.Mutex
	clients map[string]net.Conn // addr string -> connection
}

func newClientCache() *clientCache {
	return &clientCache{clients: map[string]net.Conn{}}
}

func (udp *udpProxySocket) ProxyLoop(service string, proxier *Proxier) {
	activeClients := newClientCache()
	var buffer [4096]byte // 4KiB should be enough for most whole-packets
	for {
		info, exists := proxier.getServiceInfo(service)
		if !exists {
			break
		}

		// Block until data arrives.
		// TODO: Accumulate a histogram of n or something, to fine tune the buffer size.
		n, cliAddr, err := udp.ReadFrom(buffer[0:])
		if err != nil {
			if e, ok := err.(net.Error); ok {
				if e.Temporary() {
					glog.V(1).Infof("ReadFrom had a temporary failure: %v", err)
					continue
				}
			}
			glog.Errorf("ReadFrom failed, exiting ProxyLoop: %v", err)
			break
		}
		// If this is a client we know already, reuse the connection and goroutine.
		svrConn, err := udp.getBackendConn(activeClients, cliAddr, proxier, service, info.timeout)
		if err != nil {
			continue
		}
		// TODO: It would be nice to let the goroutine handle this write, but we don't
		// really want to copy the buffer.  We could do a pool of buffers or something.
		_, err = svrConn.Write(buffer[0:n])
		if err != nil {
			if !logTimeout(err) {
				glog.Errorf("Write failed: %v", err)
				// TODO: Maybe tear down the goroutine for this client/server pair?
			}
			continue
		}
		err = svrConn.SetDeadline(time.Now().Add(info.timeout))
		if err != nil {
			glog.Errorf("SetDeadline failed: %v", err)
			continue
		}
	}
}

func (udp *udpProxySocket) getBackendConn(activeClients *clientCache, cliAddr net.Addr, proxier *Proxier, service string, timeout time.Duration) (net.Conn, error) {
	activeClients.mu.Lock()
	defer activeClients.mu.Unlock()

	svrConn, found := activeClients.clients[cliAddr.String()]
	if !found {
		// TODO: This could spin up a new goroutine to make the outbound connection,
		// and keep accepting inbound traffic.
		glog.V(2).Infof("New UDP connection from %s", cliAddr)
		var err error
		svrConn, err = tryConnect(service, cliAddr, "udp", proxier)
		if err != nil {
			return nil, err
		}
		if err = svrConn.SetDeadline(time.Now().Add(timeout)); err != nil {
			glog.Errorf("SetDeadline failed: %v", err)
			return nil, err
		}
		activeClients.clients[cliAddr.String()] = svrConn
		go func(cliAddr net.Addr, svrConn net.Conn, activeClients *clientCache, timeout time.Duration) {
			defer util.HandleCrash()
			udp.proxyClient(cliAddr, svrConn, activeClients, timeout)
		}(cliAddr, svrConn, activeClients, timeout)
	}
	return svrConn, nil
}

// This function is expected to be called as a goroutine.
// TODO: Track and log bytes copied, like TCP
func (udp *udpProxySocket) proxyClient(cliAddr net.Addr, svrConn net.Conn, activeClients *clientCache, timeout time.Duration) {
	defer svrConn.Close()
	var buffer [4096]byte
	for {
		n, err := svrConn.Read(buffer[0:])
		if err != nil {
			if !logTimeout(err) {
				glog.Errorf("Read failed: %v", err)
			}
			break
		}
		err = svrConn.SetDeadline(time.Now().Add(timeout))
		if err != nil {
			glog.Errorf("SetDeadline failed: %v", err)
			break
		}
		n, err = udp.WriteTo(buffer[0:n], cliAddr)
		if err != nil {
			if !logTimeout(err) {
				glog.Errorf("WriteTo failed: %v", err)
			}
			break
		}
	}
	activeClients.mu.Lock()
	delete(activeClients.clients, cliAddr.String())
	activeClients.mu.Unlock()
}

func logTimeout(err error) bool {
	if e, ok := err.(net.Error); ok {
		if e.Timeout() {
			glog.V(1).Infof("connection to endpoint closed due to inactivity")
			return true
		}
	}
	return false
}

func newProxySocket(protocol api.Protocol, ip net.IP, port int) (proxySocket, error) {
	host := ip.String()
	switch strings.ToUpper(string(protocol)) {
	case "TCP":
		listener, err := net.Listen("tcp", net.JoinHostPort(host, strconv.Itoa(port)))
		if err != nil {
			return nil, err
		}
		return &tcpProxySocket{listener}, nil
	case "UDP":
		addr, err := net.ResolveUDPAddr("udp", net.JoinHostPort(host, strconv.Itoa(port)))
		if err != nil {
			return nil, err
		}
		conn, err := net.ListenUDP("udp", addr)
		if err != nil {
			return nil, err
		}
		return &udpProxySocket{conn}, nil
	}
	return nil, fmt.Errorf("unknown protocol %q", protocol)
}

// Proxier is a simple proxy for TCP connections between a localhost:lport
// and services that provide the actual implementations.
type Proxier struct {
	loadBalancer LoadBalancer
	mu           sync.Mutex // protects serviceMap
	serviceMap   map[string]*serviceInfo
	listenIP     net.IP
	iptables     iptables.Interface
	hostIP       net.IP
}

// NewProxier returns a new Proxier given a LoadBalancer and an address on
// which to listen.  Because of the iptables logic, It is assumed that there
// is only a single Proxier active on a machine.
func NewProxier(loadBalancer LoadBalancer, listenIP net.IP, iptables iptables.Interface) *Proxier {
	if listenIP.Equal(localhostIPv4) || listenIP.Equal(localhostIPv6) {
		glog.Errorf("Can't proxy only on localhost - iptables can't do it")
		return nil
	}

	hostIP, err := chooseHostInterface()
	if err != nil {
		glog.Errorf("Failed to select a host interface: %v", err)
		return nil
	}

	glog.Infof("Initializing iptables")
	// Clean up old messes.  Ignore erors.
	iptablesDeleteOld(iptables)
	// Set up the iptables foundations we need.
	if err := iptablesInit(iptables); err != nil {
		glog.Errorf("Failed to initialize iptables: %v", err)
		return nil
	}
	// Flush old iptables rules (since the bound ports will be invalid after a restart).
	// When OnUpdate() is first called, the rules will be recreated.
	if err := iptablesFlush(iptables); err != nil {
		glog.Errorf("Failed to flush iptables: %v", err)
		return nil
	}
	return &Proxier{
		loadBalancer: loadBalancer,
		serviceMap:   make(map[string]*serviceInfo),
		listenIP:     listenIP,
		iptables:     iptables,
		hostIP:       hostIP,
	}
}

// The periodic interval for checking the state of things.
const syncInterval = 5 * time.Second

// SyncLoop runs periodic work.  This is expected to run as a goroutine or as the main loop of the app.  It does not return.
func (proxier *Proxier) SyncLoop() {
	for {
		select {
		case <-time.After(syncInterval):
			glog.V(2).Infof("Periodic sync")
			if err := iptablesInit(proxier.iptables); err != nil {
				glog.Errorf("Failed to ensure iptables: %v", err)
			}
			proxier.ensurePortals()
			proxier.cleanupStaleStickySessions()
		}
	}
}

// Ensure that portals exist for all services.
func (proxier *Proxier) ensurePortals() {
	proxier.mu.Lock()
	defer proxier.mu.Unlock()
	// NB: This does not remove rules that should not be present.
	for name, info := range proxier.serviceMap {
		err := proxier.openPortal(name, info)
		if err != nil {
			glog.Errorf("Failed to ensure portal for %q: %v", name, err)
		}
	}
}

// clean up any stale sticky session records in the hash map.
func (proxier *Proxier) cleanupStaleStickySessions() {
	for name, info := range proxier.serviceMap {
		if info.sessionAffinityType != api.AffinityTypeNone {
			proxier.loadBalancer.CleanupStaleStickySessions(name)
		}
	}
}

// This assumes proxier.mu is not locked.
func (proxier *Proxier) stopProxy(service string, info *serviceInfo) error {
	proxier.mu.Lock()
	defer proxier.mu.Unlock()
	return proxier.stopProxyInternal(service, info)
}

// This assumes proxier.mu is locked.
func (proxier *Proxier) stopProxyInternal(service string, info *serviceInfo) error {
	delete(proxier.serviceMap, service)
	return info.socket.Close()
}

func (proxier *Proxier) getServiceInfo(service string) (*serviceInfo, bool) {
	proxier.mu.Lock()
	defer proxier.mu.Unlock()
	info, ok := proxier.serviceMap[service]
	return info, ok
}

func (proxier *Proxier) setServiceInfo(service string, info *serviceInfo) {
	proxier.mu.Lock()
	defer proxier.mu.Unlock()
	proxier.serviceMap[service] = info
}

// addServiceOnPort starts listening for a new service, returning the serviceInfo.
// Pass proxyPort=0 to allocate a random port. The timeout only applies to UDP
// connections, for now.
func (proxier *Proxier) addServiceOnPort(service string, protocol api.Protocol, proxyPort int, timeout time.Duration) (*serviceInfo, error) {
	sock, err := newProxySocket(protocol, proxier.listenIP, proxyPort)
	if err != nil {
		return nil, err
	}
	_, portStr, err := net.SplitHostPort(sock.Addr().String())
	if err != nil {
		sock.Close()
		return nil, err
	}
	portNum, err := strconv.Atoi(portStr)
	if err != nil {
		sock.Close()
		return nil, err
	}
	si := &serviceInfo{
		proxyPort:           portNum,
		protocol:            protocol,
		socket:              sock,
		timeout:             timeout,
		sessionAffinityType: api.AffinityTypeNone,
		stickyMaxAgeMinutes: 180,
	}
	proxier.setServiceInfo(service, si)

	glog.V(1).Infof("Proxying for service %q on %s port %d", service, protocol, portNum)
	go func(service string, proxier *Proxier) {
		defer util.HandleCrash()
		sock.ProxyLoop(service, proxier)
	}(service, proxier)

	return si, nil
}

// How long we leave idle UDP connections open.
const udpIdleTimeout = 1 * time.Minute

// OnUpdate manages the active set of service proxies.
// Active service proxies are reinitialized if found in the update set or
// shutdown if missing from the update set.
func (proxier *Proxier) OnUpdate(services []api.Service) {
	glog.V(4).Infof("Received update notice: %+v", services)
	activeServices := util.StringSet{}
	for _, service := range services {
		activeServices.Insert(service.Name)
		info, exists := proxier.getServiceInfo(service.Name)
		serviceIP := net.ParseIP(service.Spec.PortalIP)
		// TODO: check health of the socket?  What if ProxyLoop exited?
		if exists && info.portalPort == service.Spec.Port && info.portalIP.Equal(serviceIP) {
			continue
		}
		if exists && (info.portalPort != service.Spec.Port || !info.portalIP.Equal(serviceIP) || !ipsEqual(service.Spec.PublicIPs, info.publicIP)) {
			glog.V(4).Infof("Something changed for service %q: stopping it", service.Name)
			err := proxier.closePortal(service.Name, info)
			if err != nil {
				glog.Errorf("Failed to close portal for %q: %v", service.Name, err)
			}
			err = proxier.stopProxy(service.Name, info)
			if err != nil {
				glog.Errorf("Failed to stop service %q: %v", service.Name, err)
			}
		}
		glog.V(1).Infof("Adding new service %q at %s:%d/%s (local :%d)", service.Name, serviceIP, service.Spec.Port, service.Spec.Protocol, service.Spec.ProxyPort)
		info, err := proxier.addServiceOnPort(service.Name, service.Spec.Protocol, service.Spec.ProxyPort, udpIdleTimeout)
		if err != nil {
			glog.Errorf("Failed to start proxy for %q: %v", service.Name, err)
			continue
		}
		info.portalIP = serviceIP
		info.portalPort = service.Spec.Port
		info.publicIP = service.Spec.PublicIPs
		info.sessionAffinityType = service.Spec.SessionAffinity
		// TODO: paramaterize this in the types api file as an attribute of sticky session.   For now it's hardcoded to 3 hours.
		info.stickyMaxAgeMinutes = 180
		glog.V(4).Infof("info: %+v", info)

		err = proxier.openPortal(service.Name, info)
		if err != nil {
			glog.Errorf("Failed to open portal for %q: %v", service.Name, err)
		}
		proxier.loadBalancer.NewService(service.Name, info.sessionAffinityType, info.stickyMaxAgeMinutes)
	}
	proxier.mu.Lock()
	defer proxier.mu.Unlock()
	for name, info := range proxier.serviceMap {
		if !activeServices.Has(name) {
			glog.V(1).Infof("Stopping service %q", name)
			err := proxier.closePortal(name, info)
			if err != nil {
				glog.Errorf("Failed to close portal for %q: %v", name, err)
			}
			err = proxier.stopProxyInternal(name, info)
			if err != nil {
				glog.Errorf("Failed to stop service %q: %v", name, err)
			}
		}
	}
}

func ipsEqual(lhs, rhs []string) bool {
	if len(lhs) != len(rhs) {
		return false
	}
	for i := range lhs {
		if lhs[i] != rhs[i] {
			return false
		}
	}
	return true
}

func (proxier *Proxier) openPortal(service string, info *serviceInfo) error {
	err := proxier.openOnePortal(info.portalIP, info.portalPort, info.protocol, proxier.listenIP, info.proxyPort, service)
	if err != nil {
		return err
	}
	for _, publicIP := range info.publicIP {
		err = proxier.openOnePortal(net.ParseIP(publicIP), info.portalPort, info.protocol, proxier.listenIP, info.proxyPort, service)
		if err != nil {
			return err
		}
	}
	return nil
}

func (proxier *Proxier) openOnePortal(portalIP net.IP, portalPort int, protocol api.Protocol, proxyIP net.IP, proxyPort int, name string) error {
	// Handle traffic from containers.
	args := proxier.iptablesContainerPortalArgs(portalIP, portalPort, protocol, proxyIP, proxyPort, name)
	existed, err := proxier.iptables.EnsureRule(iptables.TableNAT, iptablesContainerPortalChain, args...)
	if err != nil {
		glog.Errorf("Failed to install iptables %s rule for service %q", iptablesContainerPortalChain, name)
		return err
	}
	if !existed {
		glog.Infof("Opened iptables from-containers portal for service %q on %s %s:%d", name, protocol, portalIP, portalPort)
	}

	// Handle traffic from the host.
	args = proxier.iptablesHostPortalArgs(portalIP, portalPort, protocol, proxyIP, proxyPort, name)
	existed, err = proxier.iptables.EnsureRule(iptables.TableNAT, iptablesHostPortalChain, args...)
	if err != nil {
		glog.Errorf("Failed to install iptables %s rule for service %q", iptablesHostPortalChain, name)
		return err
	}
	if !existed {
		glog.Infof("Opened iptables from-host portal for service %q on %s %s:%d", name, protocol, portalIP, portalPort)
	}
	return nil
}

func (proxier *Proxier) closePortal(service string, info *serviceInfo) error {
	// Collect errors and report them all at the end.
	el := proxier.closeOnePortal(info.portalIP, info.portalPort, info.protocol, proxier.listenIP, info.proxyPort, service)
	for _, publicIP := range info.publicIP {
		el = append(el, proxier.closeOnePortal(net.ParseIP(publicIP), info.portalPort, info.protocol, proxier.listenIP, info.proxyPort, service)...)
	}
	if len(el) == 0 {
		glog.Infof("Closed iptables portals for service %q", service)
	} else {
		glog.Errorf("Some errors closing iptables portals for service %q", service)
	}
	return errors.NewAggregate(el)
}

func (proxier *Proxier) closeOnePortal(portalIP net.IP, portalPort int, protocol api.Protocol, proxyIP net.IP, proxyPort int, name string) []error {
	el := []error{}

	// Handle traffic from containers.
	args := proxier.iptablesContainerPortalArgs(portalIP, portalPort, protocol, proxyIP, proxyPort, name)
	if err := proxier.iptables.DeleteRule(iptables.TableNAT, iptablesContainerPortalChain, args...); err != nil {
		glog.Errorf("Failed to delete iptables %s rule for service %q", iptablesContainerPortalChain, name)
		el = append(el, err)
	}

	// Handle traffic from the host.
	args = proxier.iptablesHostPortalArgs(portalIP, portalPort, protocol, proxyIP, proxyPort, name)
	if err := proxier.iptables.DeleteRule(iptables.TableNAT, iptablesHostPortalChain, args...); err != nil {
		glog.Errorf("Failed to delete iptables %s rule for service %q", iptablesHostPortalChain, name)
		el = append(el, err)
	}

	return el
}

// See comments in the *PortalArgs() functions for some details about why we
// use two chains.
var iptablesContainerPortalChain iptables.Chain = "KUBE-PORTALS-CONTAINER"
var iptablesHostPortalChain iptables.Chain = "KUBE-PORTALS-HOST"
var iptablesOldPortalChain iptables.Chain = "KUBE-PROXY"

// Ensure that the iptables infrastructure we use is set up.  This can safely be called periodically.
func iptablesInit(ipt iptables.Interface) error {
	// TODO: There is almost certainly room for optimization here.  E.g. If
	// we knew the portal_net CIDR we could fast-track outbound packets not
	// destined for a service. There's probably more, help wanted.
	if _, err := ipt.EnsureChain(iptables.TableNAT, iptablesContainerPortalChain); err != nil {
		return err
	}
	if _, err := ipt.EnsureRule(iptables.TableNAT, iptables.ChainPrerouting, "-j", string(iptablesContainerPortalChain)); err != nil {
		return err
	}
	if _, err := ipt.EnsureChain(iptables.TableNAT, iptablesHostPortalChain); err != nil {
		return err
	}
	if _, err := ipt.EnsureRule(iptables.TableNAT, iptables.ChainOutput, "-j", string(iptablesHostPortalChain)); err != nil {
		return err
	}
	return nil
}

func iptablesDeleteOld(ipt iptables.Interface) {
	// DEPRECATED: The iptablesOldPortalChain is from when we had a single chain
	// for all rules.  We'll unilaterally delete it here.  We will remove this
	// code at some future date (before 1.0).
	ipt.DeleteRule(iptables.TableNAT, iptables.ChainPrerouting, "-j", string(iptablesOldPortalChain))
	ipt.DeleteRule(iptables.TableNAT, iptables.ChainOutput, "-j", string(iptablesOldPortalChain))
	ipt.FlushChain(iptables.TableNAT, iptablesOldPortalChain)
	ipt.DeleteChain(iptables.TableNAT, iptablesOldPortalChain)
}

// Flush all of our custom iptables rules.
func iptablesFlush(ipt iptables.Interface) error {
	el := []error{}
	if err := ipt.FlushChain(iptables.TableNAT, iptablesContainerPortalChain); err != nil {
		el = append(el, err)
	}
	if err := ipt.FlushChain(iptables.TableNAT, iptablesHostPortalChain); err != nil {
		el = append(el, err)
	}
	if len(el) != 0 {
		glog.Errorf("Some errors flushing old iptables portals: %v", el)
	}
	return errors.NewAggregate(el)
}

// Used below.
var zeroIPv4 = net.ParseIP("0.0.0.0")
var localhostIPv4 = net.ParseIP("127.0.0.1")

var zeroIPv6 = net.ParseIP("::0")
var localhostIPv6 = net.ParseIP("::1")

// Build a slice of iptables args that are common to from-container and from-host portal rules.
func iptablesCommonPortalArgs(destIP net.IP, destPort int, protocol api.Protocol, service string) []string {
	// This list needs to include all fields as they are eventually spit out
	// by iptables-save.  This is because some systems do not support the
	// 'iptables -C' arg, and so fall back on parsing iptables-save output.
	// If this does not match, it will not pass the check.  For example:
	// adding the /32 on the destination IP arg is not strictly required,
	// but causes this list to not match the final iptables-save output.
	// This is fragile and I hope one day we can stop supporting such old
	// iptables versions.
	args := []string{
		"-m", "comment",
		"--comment", service,
		"-p", strings.ToLower(string(protocol)),
		"-m", strings.ToLower(string(protocol)),
		"-d", fmt.Sprintf("%s/32", destIP.String()),
		"--dport", fmt.Sprintf("%d", destPort),
	}
	return args
}

// Build a slice of iptables args for a from-container portal rule.
func (proxier *Proxier) iptablesContainerPortalArgs(destIP net.IP, destPort int, protocol api.Protocol, proxyIP net.IP, proxyPort int, service string) []string {
	args := iptablesCommonPortalArgs(destIP, destPort, protocol, service)

	// This is tricky.
	//
	// If the proxy is bound (see Proxier.listenIP) to 0.0.0.0 ("any
	// interface") we want to use REDIRECT, which sends traffic to the
	// "primary address of the incoming interface" which means the container
	// bridge, if there is one.  When the response comes, it comes from that
	// same interface, so the NAT matches and the response packet is
	// correct.  This matters for UDP, since there is no per-connection port
	// number.
	//
	// The alternative would be to use DNAT, except that it doesn't work
	// (empirically):
	//   * DNAT to 127.0.0.1 = Packets just disappear - this seems to be a
	//     well-known limitation of iptables.
	//   * DNAT to eth0's IP = Response packets come from the bridge, which
	//     breaks the NAT, and makes things like DNS not accept them.  If
	//     this could be resolved, it would simplify all of this code.
	//
	// If the proxy is bound to a specific IP, then we have to use DNAT to
	// that IP.  Unlike the previous case, this works because the proxy is
	// ONLY listening on that IP, not the bridge.
	//
	// Why would anyone bind to an address that is not inclusive of
	// localhost?  Apparently some cloud environments have their public IP
	// exposed as a real network interface AND do not have firewalling.  We
	// don't want to expose everything out to the world.
	//
	// Unfortunately, I don't know of any way to listen on some (N > 1)
	// interfaces but not ALL interfaces, short of doing it manually, and
	// this is simpler than that.
	//
	// If the proxy is bound to localhost only, all of this is broken.  Not
	// allowed.
	if proxyIP.Equal(zeroIPv4) || proxyIP.Equal(zeroIPv6) {
		// TODO: Can we REDIRECT with IPv6?
		args = append(args, "-j", "REDIRECT", "--to-ports", fmt.Sprintf("%d", proxyPort))
	} else {
		// TODO: Can we DNAT with IPv6?
		args = append(args, "-j", "DNAT", "--to-destination", net.JoinHostPort(proxyIP.String(), strconv.Itoa(proxyPort)))
	}
	return args
}

// Build a slice of iptables args for a from-host portal rule.
func (proxier *Proxier) iptablesHostPortalArgs(destIP net.IP, destPort int, protocol api.Protocol, proxyIP net.IP, proxyPort int, service string) []string {
	args := iptablesCommonPortalArgs(destIP, destPort, protocol, service)

	// This is tricky.
	//
	// If the proxy is bound (see Proxier.listenIP) to 0.0.0.0 ("any
	// interface") we want to do the same as from-container traffic and use
	// REDIRECT.  Except that it doesn't work (empirically).  REDIRECT on
	// localpackets sends the traffic to localhost (special case, but it is
	// documented) but the response comes from the eth0 IP (not sure why,
	// truthfully), which makes DNS unhappy.
	//
	// So we have to use DNAT.  DNAT to 127.0.0.1 can't work for the same
	// reason.
	//
	// So we do our best to find an interface that is not a loopback and
	// DNAT to that.  This works (again, empirically).
	//
	// If the proxy is bound to a specific IP, then we have to use DNAT to
	// that IP.  Unlike the previous case, this works because the proxy is
	// ONLY listening on that IP, not the bridge.
	//
	// If the proxy is bound to localhost only, this should work, but we
	// don't allow it for now.
	if proxyIP.Equal(zeroIPv4) || proxyIP.Equal(zeroIPv6) {
		proxyIP = proxier.hostIP
	}
	// TODO: Can we DNAT with IPv6?
	args = append(args, "-j", "DNAT", "--to-destination", net.JoinHostPort(proxyIP.String(), strconv.Itoa(proxyPort)))
	return args
}

func chooseHostInterface() (net.IP, error) {
	intfs, err := net.Interfaces()
	if err != nil {
		return nil, err
	}
	i := 0
	for i = range intfs {
		if flagsSet(intfs[i].Flags, net.FlagUp) && flagsClear(intfs[i].Flags, net.FlagLoopback|net.FlagPointToPoint) {
			addrs, err := intfs[i].Addrs()
			if err != nil {
				return nil, err
			}
			if len(addrs) > 0 {
				// This interface should suffice.
				break
			}
		}
	}
	if i == len(intfs) {
		return nil, err
	}
	glog.V(2).Infof("Choosing interface %s for from-host portals", intfs[i].Name)
	addrs, err := intfs[i].Addrs()
	if err != nil {
		return nil, err
	}
	glog.V(2).Infof("Interface %s = %s", intfs[i].Name, addrs[0].String())
	ip, _, err := net.ParseCIDR(addrs[0].String())
	if err != nil {
		return nil, err
	}
	return ip, nil
}

func flagsSet(flags net.Flags, test net.Flags) bool {
	return flags&test != 0
}

func flagsClear(flags net.Flags, test net.Flags) bool {
	return flags&test == 0
}