net/netns, wgengine/router: support Linux machines that don't have 'ip rule'.

We'll use SO_BINDTODEVICE instead of fancy policy routing. This has
some limitations: for example, we will route all traffic through the
interface that has the main "default" (0.0.0.0/0) route, so machines
that have multiple physical interfaces might have to go through DERP to
get to some peers. But machines with multiple physical interfaces are
very likely to have policy routing (ip rule) support anyway.

So far, the only OS I know of that needs this feature is ChromeOS
(crostini). Fixes #245.

Signed-off-by: Avery Pennarun <apenwarr@tailscale.com>

net/netns/netns_linux.go

@@ -7,9 +7,14 @@
 package netns
 
 import (
+	"errors"
 	"fmt"
+	"io/ioutil"
 	"os"
+	"os/exec"
 	"path/filepath"
+	"strings"
+	"sync"
 	"syscall"
 
 	"golang.org/x/sys/unix"
@@ -23,44 +28,126 @@ import (
 // wgengine/router/router_linux.go.
 const tailscaleBypassMark = 0x20000
 
+// checkIPRule runs the ipRuleAvailable check exactly once.
+var checkIPRule sync.Once
+
+// ipRuleAvailable is true if and only if the 'ip rule' command works.
+// If it doesn't, we have to use SO_BINDTODEVICE on our sockets instead.
+var ipRuleAvailable bool
+
+// defaultRouteInterface returns the name of the network interface that owns
+// the default route, not including any tailscale interfaces. We only use
+// this in SO_BINDTODEVICE mode.
+func defaultRouteInterface() (string, error) {
+	b, err := ioutil.ReadFile("/proc/net/route")
+	if err != nil {
+		return "", err
+	}
+
+	for _, line := range strings.Split(string(b), "\n")[1:] {
+		fields := strings.Fields(line)
+		ifc := fields[0]
+		ip := fields[1]
+		netmask := fields[7]
+
+		if strings.HasPrefix(ifc, "tailscale") ||
+			strings.HasPrefix(ifc, "wg") {
+			continue
+		}
+		if ip == "00000000" && netmask == "00000000" {
+			// default route
+			return ifc, nil // interface name
+		}
+	}
+
+	return "", errors.New("no default routes found")
+}
+
+// ignoreErrors returns true if we should ignore setsocketopt errors in
+// this instance.
+func ignoreErrors() bool {
+	if os.Getuid() != 0 {
+		// only root can manipulate these socket flags
+		return true
+	}
+
+	// TODO(apenwarr): this snooping around in the args is way too magic.
+	//  It would be better to explicitly activate, or not, this dialer
+	//  by passing it from the toplevel program.
+	v, _ := os.Executable()
+	switch filepath.Base(v) {
+	case "tailscale":
+		for _, arg := range os.Args {
+			if arg == "netcheck" {
+				return true
+			}
+		}
+	case "tailscaled":
+		for _, arg := range os.Args {
+			if arg == "-fake" || arg == "--fake" {
+				return true
+			}
+		}
+	}
+
+	return false
+}
+
 // control marks c as necessary to dial in a separate network namespace.
 //
 // It's intentionally the same signature as net.Dialer.Control
 // and net.ListenConfig.Control.
 func control(network, address string, c syscall.RawConn) error {
+	checkIPRule.Do(func() {
+		_, err := exec.Command("ip", "rule").Output()
+		ipRuleAvailable = (err == nil)
+	})
+
 	if skipPrivileged.Get() {
 		// We can't set socket marks without CAP_NET_ADMIN on linux,
 		// skip as requested.
 		return nil
 	}
 
-	var controlErr error
-	err := c.Control(func(fd uintptr) {
-		controlErr = unix.SetsockoptInt(int(fd), unix.SOL_SOCKET, unix.SO_MARK, tailscaleBypassMark)
-	})
-	// Before returning some fatal error, skip it in some cases.
-	if (err != nil || controlErr != nil) && os.Getuid() != 0 {
-		v, _ := os.Executable()
-		switch filepath.Base(v) {
-		case "tailscale":
-			for _, arg := range os.Args {
-				if arg == "netcheck" {
-					return nil
-				}
-			}
-		case "tailscaled":
-			for _, arg := range os.Args {
-				if arg == "-fake" || arg == "--fake" {
-					return nil
-				}
+	if ipRuleAvailable {
+		var controlErr error
+		err := c.Control(func(fd uintptr) {
+			controlErr = unix.SetsockoptInt(int(fd),
+				unix.SOL_SOCKET, unix.SO_MARK,
+				tailscaleBypassMark)
+		})
+		if (err != nil || controlErr != nil) && ignoreErrors() {
+			return nil
+		}
+		if err != nil {
+			return fmt.Errorf("setting socket mark1: %w", err)
+		}
+		if controlErr != nil {
+			return fmt.Errorf("setting socket mark2: %w", controlErr)
+		}
+	} else {
+		var controlErr error
+		err := c.Control(func(fd uintptr) {
+			ifc, err := defaultRouteInterface()
+			if err != nil {
+				// Make sure we bind to *some* interface,
+				// or we could get a routing loop.
+				// "lo" is always wrong, but if we don't have
+				// a default route anyway, it doesn't matter.
+				ifc = "lo"
 			}
+			controlErr = unix.SetsockoptString(int(fd),
+				unix.SOL_SOCKET, unix.SO_BINDTODEVICE, ifc)
+		})
+		if (err != nil || controlErr != nil) && ignoreErrors() {
+			return nil
+		}
+		if err != nil {
+			return fmt.Errorf("setting SO_BINDTODEVICE 1: %w", err)
+		}
+		if controlErr != nil {
+			return fmt.Errorf("setting SO_BINDTODEVICE 2: %w", controlErr)
 		}
-	}
-	if err != nil {
-		return fmt.Errorf("setting socket mark: %w", err)
-	}
-	if controlErr != nil {
-		return fmt.Errorf("setting socket mark: %w", controlErr)
 	}
 	return nil
 }

wgengine/router/router_linux.go

@@ -70,6 +70,7 @@ type netfilterRunner interface {
 
 type linuxRouter struct {
 	logf             func(fmt string, args ...interface{})
+	ipRuleAvailable  bool
 	tunname          string
 	addrs            map[netaddr.IPPrefix]bool
 	routes           map[netaddr.IPPrefix]bool
@@ -96,12 +97,16 @@ func newUserspaceRouter(logf logger.Logf, _ *device.Device, tunDev tun.Device) (
 }
 
 func newUserspaceRouterAdvanced(logf logger.Logf, tunname string, netfilter netfilterRunner, cmd commandRunner) (Router, error) {
+	_, err := exec.Command("ip", "rule").Output()
+	ipRuleAvailable := (err == nil)
+
 	return &linuxRouter{
-		logf:          logf,
-		tunname:       tunname,
-		netfilterMode: NetfilterOff,
-		ipt4:          netfilter,
-		cmd:           cmd,
+		logf:            logf,
+		ipRuleAvailable: ipRuleAvailable,
+		tunname:         tunname,
+		netfilterMode:   NetfilterOff,
+		ipt4:            netfilter,
+		cmd:             cmd,
 	}, nil
 }
 
@@ -109,10 +114,10 @@ func (r *linuxRouter) Up() error {
 	if err := r.delLegacyNetfilter(); err != nil {
 		return err
 	}
-	if err := r.setNetfilterMode(NetfilterOff); err != nil {
+	if err := r.addIPRules(); err != nil {
 		return err
 	}
-	if err := r.addBypassRule(); err != nil {
+	if err := r.setNetfilterMode(NetfilterOff); err != nil {
 		return err
 	}
 	if err := r.upInterface(); err != nil {
@@ -126,7 +131,7 @@ func (r *linuxRouter) down() error {
 	if err := r.downInterface(); err != nil {
 		return err
 	}
-	if err := r.delBypassRule(); err != nil {
+	if err := r.delIPRules(); err != nil {
 		return err
 	}
 	if err := r.setNetfilterMode(NetfilterOff); err != nil {
@@ -477,24 +482,30 @@ func (r *linuxRouter) delLoopbackRule(addr netaddr.IP) error {
 // interface. Fails if the route already exists, or if adding the
 // route fails.
 func (r *linuxRouter) addRoute(cidr netaddr.IPPrefix) error {
-	return r.cmd.run(
+	args := []string{
 		"ip", "route", "add",
 		normalizeCIDR(cidr),
 		"dev", r.tunname,
-		"table", "88",
-	)
+	}
+	if r.ipRuleAvailable {
+		args = append(args, "table", "88")
+	}
+	return r.cmd.run(args...)
 }
 
 // delRoute removes the route for cidr pointing to the tunnel
 // interface. Fails if the route doesn't exist, or if removing the
 // route fails.
 func (r *linuxRouter) delRoute(cidr netaddr.IPPrefix) error {
-	return r.cmd.run(
+	args := []string{
 		"ip", "route", "del",
 		normalizeCIDR(cidr),
 		"dev", r.tunname,
-		"table", "88",
-	)
+	}
+	if r.ipRuleAvailable {
+		args = append(args, "table", "88")
+	}
+	return r.cmd.run(args...)
 }
 
 // addSubnetRule adds a netfilter rule that allows traffic to flow
@@ -540,13 +551,17 @@ func (r *linuxRouter) downInterface() error {
 	return r.cmd.run("ip", "link", "set", "dev", r.tunname, "down")
 }
 
-// addBypassRule adds the policy routing rule that avoids tailscaled
+// addIPRules adds the policy routing rule that avoids tailscaled
 // routing loops. If the rule exists and appears to be a
 // tailscale-managed rule, it is gracefully replaced.
-func (r *linuxRouter) addBypassRule() error {
+func (r *linuxRouter) addIPRules() error {
+	if !r.ipRuleAvailable {
+		return nil
+	}
+
 	// Clear out old rules. After that, any error adding a rule is fatal,
 	// because there should be no reason we add a duplicate.
-	if err := r.delBypassRule(); err != nil {
+	if err := r.delIPRules(); err != nil {
 		return err
 	}
 
@@ -611,7 +626,11 @@ func (r *linuxRouter) addBypassRule() error {
 
 // delBypassrule removes the policy routing rules that avoid
 // tailscaled routing loops, if it exists.
-func (r *linuxRouter) delBypassRule() error {
+func (r *linuxRouter) delIPRules() error {
+	if !r.ipRuleAvailable {
+		return nil
+	}
+
 	// Error codes: 'ip rule' returns error code 2 if the rule is a
 	// duplicate (add) or not found (del). It returns a different code
 	// for syntax errors. This is also true of busybox.