tailscale/wgengine/router/router_linux.go

// Copyright (c) Tailscale Inc & AUTHORS
// SPDX-License-Identifier: BSD-3-Clause

package router

import (
	"errors"
	"fmt"
	"net"
	"net/netip"
	"os"
	"os/exec"
	"strconv"
	"strings"
	"sync/atomic"
	"syscall"
	"time"

	"github.com/tailscale/netlink"
	"github.com/tailscale/wireguard-go/tun"
	"go4.org/netipx"
	"golang.org/x/sys/unix"
	"golang.org/x/time/rate"
	"tailscale.com/envknob"
	"tailscale.com/net/netmon"
	"tailscale.com/types/logger"
	"tailscale.com/types/preftype"
	"tailscale.com/util/linuxfw"
	"tailscale.com/util/multierr"
	"tailscale.com/version/distro"
)

const (
	netfilterOff      = preftype.NetfilterOff
	netfilterNoDivert = preftype.NetfilterNoDivert
	netfilterOn       = preftype.NetfilterOn
)

type linuxRouter struct {
	closed           atomic.Bool
	logf             func(fmt string, args ...any)
	tunname          string
	netMon           *netmon.Monitor
	unregNetMon      func()
	addrs            map[netip.Prefix]bool
	routes           map[netip.Prefix]bool
	localRoutes      map[netip.Prefix]bool
	snatSubnetRoutes bool
	netfilterMode    preftype.NetfilterMode
	netfilterKind    string

	// ruleRestorePending is whether a timer has been started to
	// restore deleted ip rules.
	ruleRestorePending atomic.Bool
	ipRuleFixLimiter   *rate.Limiter

	// Various feature checks for the network stack.
	ipRuleAvailable bool // whether kernel was built with IP_MULTIPLE_TABLES
	v6Available     bool // whether the kernel supports IPv6
	fwmaskWorks     bool // whether we can use 'ip rule...fwmark <mark>/<mask>'

	// ipPolicyPrefBase is the base priority at which ip rules are installed.
	ipPolicyPrefBase int

	cmd commandRunner
	nfr linuxfw.NetfilterRunner

	magicsockPortV4 uint16
	magicsockPortV6 uint16
}

func newUserspaceRouter(logf logger.Logf, tunDev tun.Device, netMon *netmon.Monitor) (Router, error) {
	tunname, err := tunDev.Name()
	if err != nil {
		return nil, err
	}

	cmd := osCommandRunner{
		ambientCapNetAdmin: useAmbientCaps(),
	}

	return newUserspaceRouterAdvanced(logf, tunname, netMon, cmd)
}

func newUserspaceRouterAdvanced(logf logger.Logf, tunname string, netMon *netmon.Monitor, cmd commandRunner) (Router, error) {
	r := &linuxRouter{
		logf:          logf,
		tunname:       tunname,
		netfilterMode: netfilterOff,
		netMon:        netMon,

		cmd: cmd,

		ipRuleFixLimiter: rate.NewLimiter(rate.Every(5*time.Second), 10),
		ipPolicyPrefBase: 5200,
	}
	if r.useIPCommand() {
		r.ipRuleAvailable = (cmd.run("ip", "rule") == nil)
	} else {
		if rules, err := netlink.RuleList(netlink.FAMILY_V4); err != nil {
			r.logf("error querying IP rules (does kernel have IP_MULTIPLE_TABLES?): %v", err)
			r.logf("warning: running without policy routing")
		} else {
			r.logf("[v1] policy routing available; found %d rules", len(rules))
			r.ipRuleAvailable = true
		}
	}

	// To be a good denizen of the 4-byte 'fwmark' bitspace on every packet, we try to
	// only use the third byte. However, support for masking to part of the fwmark bitspace
	// was only added to busybox in 1.33.0. As such, we want to detect older versions and
	// not issue such a stanza.
	var err error
	if r.fwmaskWorks, err = ipCmdSupportsFwmask(); err != nil {
		r.logf("failed to determine ip command fwmask support: %v", err)
	}
	if r.fwmaskWorks {
		r.logf("[v1] ip command supports fwmark masks")
	} else {
		r.logf("[v1] ip command does NOT support fwmark masks")
	}

	// A common installation of OpenWRT involves use of the 'mwan3' package.
	// This package installs ip-tables rules like:
	//  -A mwan3_fallback_policy -m mark --mark 0x0/0x3f00 -j MARK --set-xmark 0x100/0x3f00
	//
	// which coupled with an ip rule:
	//  2001: from all fwmark 0x100/0x3f00 lookup 1
	//
	// has the effect of gobbling tailscale packets, because tailscale by default installs
	// its policy routing rules at priority 52xx.
	//
	// As such, if we are running on openWRT, detect a mwan3 config, AND detect a rule
	// with a preference 2001 (corresponding to the first interface wman3 manages), we
	// shift the priority of our policies to 13xx. This effectively puts us between mwan3's
	// permit-by-src-ip rules and mwan3 lookup of its own routing table which would drop
	// the packet.
	isMWAN3, err := checkOpenWRTUsingMWAN3()
	if err != nil {
		r.logf("error checking mwan3 installation: %v", err)
	} else if isMWAN3 {
		r.ipPolicyPrefBase = 1300
		r.logf("mwan3 on openWRT detected, switching policy base priority to 1300")
	}

	r.v6Available = linuxfw.CheckIPv6(r.logf) == nil

	r.fixupWSLMTU()

	return r, nil
}

// ipCmdSupportsFwmask returns true if the system 'ip' binary supports using a
// fwmark stanza with a mask specified. To our knowledge, everything except busybox
// pre-1.33 supports this.
func ipCmdSupportsFwmask() (bool, error) {
	ipPath, err := exec.LookPath("ip")
	if err != nil {
		return false, fmt.Errorf("lookpath: %v", err)
	}
	stat, err := os.Lstat(ipPath)
	if err != nil {
		return false, fmt.Errorf("lstat: %v", err)
	}
	if stat.Mode()&os.ModeSymlink == 0 {
		// Not a symlink, so can't be busybox. Must be regular ip utility.
		return true, nil
	}

	linkDest, err := os.Readlink(ipPath)
	if err != nil {
		return false, err
	}
	if !strings.Contains(strings.ToLower(linkDest), "busybox") {
		// Not busybox, presumably supports fwmark masks.
		return true, nil
	}

	// If we got this far, the ip utility is a busybox version with an
	// unknown version.
	// We run `ip --version` and look for the busybox banner (which
	// is a stable 'BusyBox vX.Y.Z (<builddate>)' string) to determine
	// the version.
	out, err := exec.Command("ip", "--version").CombinedOutput()
	if err != nil {
		return false, err
	}
	major, minor, _, err := busyboxParseVersion(string(out))
	if err != nil {
		return false, nil
	}

	// Support for masks added in 1.33.0.
	switch {
	case major > 1:
		return true, nil
	case major == 1 && minor >= 33:
		return true, nil
	default:
		return false, nil
	}
}

func busyboxParseVersion(output string) (major, minor, patch int, err error) {
	bannerStart := strings.Index(output, "BusyBox v")
	if bannerStart < 0 {
		return 0, 0, 0, errors.New("missing BusyBox banner")
	}
	bannerEnd := bannerStart + len("BusyBox v")

	end := strings.Index(output[bannerEnd:], " ")
	if end < 0 {
		return 0, 0, 0, errors.New("missing end delimiter")
	}

	elements := strings.Split(output[bannerEnd:bannerEnd+end], ".")
	if len(elements) < 3 {
		return 0, 0, 0, fmt.Errorf("expected 3 version elements, got %d", len(elements))
	}

	if major, err = strconv.Atoi(elements[0]); err != nil {
		return 0, 0, 0, fmt.Errorf("parsing major: %v", err)
	}
	if minor, err = strconv.Atoi(elements[1]); err != nil {
		return 0, 0, 0, fmt.Errorf("parsing minor: %v", err)
	}
	if patch, err = strconv.Atoi(elements[2]); err != nil {
		return 0, 0, 0, fmt.Errorf("parsing patch: %v", err)
	}
	return major, minor, patch, nil
}

func useAmbientCaps() bool {
	if distro.Get() != distro.Synology {
		return false
	}
	return distro.DSMVersion() >= 7
}

var forceIPCommand = envknob.RegisterBool("TS_DEBUG_USE_IP_COMMAND")

// useIPCommand reports whether r should use the "ip" command (or its
// fake commandRunner for tests) instead of netlink.
func (r *linuxRouter) useIPCommand() bool {
	if r.cmd == nil {
		panic("invalid init")
	}
	if forceIPCommand() {
		return true
	}
	// In the future we might need to fall back to using the "ip"
	// command if, say, netlink is blocked somewhere but the ip
	// command is allowed to use netlink. For now we only use the ip
	// command runner in tests.
	_, ok := r.cmd.(osCommandRunner)
	return !ok
}

// onIPRuleDeleted is the callback from the network monitor for when an IP
// policy rule is deleted. See Issue 1591.
//
// If an ip rule is deleted (with pref number 52xx, as Tailscale sets), then
// set a timer to restore our rules, in case they were deleted. The timer lets
// us do one fixup in response to a batch of rule deletes. It also lets us
// delay arbitrarily to prevent a high-speed fight over the rule between
// competing processes. (Although empirically, systemd doesn't fight us
// like that... yet.)
//
// Note that we don't care about the table number. We don't strictly even care
// about the priority number. We could just do this in response to any netlink
// change. Filtering by known priority ranges cuts back on some logspam.
func (r *linuxRouter) onIPRuleDeleted(table uint8, priority uint32) {
	if int(priority) < r.ipPolicyPrefBase || int(priority) >= (r.ipPolicyPrefBase+100) {
		// Not our rule.
		return
	}
	if r.ruleRestorePending.Swap(true) {
		// Another timer is already pending.
		return
	}
	rr := r.ipRuleFixLimiter.Reserve()
	if !rr.OK() {
		r.ruleRestorePending.Swap(false)
		return
	}
	time.AfterFunc(rr.Delay()+250*time.Millisecond, func() {
		if r.ruleRestorePending.Swap(false) && !r.closed.Load() {
			r.logf("somebody (likely systemd-networkd) deleted ip rules; restoring Tailscale's")
			r.justAddIPRules()
		}
	})
}

func (r *linuxRouter) Up() error {
	if r.unregNetMon == nil && r.netMon != nil {
		r.unregNetMon = r.netMon.RegisterRuleDeleteCallback(r.onIPRuleDeleted)
	}
	if err := r.setNetfilterMode(netfilterOff); err != nil {
		return fmt.Errorf("setting netfilter mode: %w", err)
	}
	if err := r.addIPRules(); err != nil {
		return fmt.Errorf("adding IP rules: %w", err)
	}
	if err := r.upInterface(); err != nil {
		return fmt.Errorf("bringing interface up: %w", err)
	}

	return nil
}

func (r *linuxRouter) Close() error {
	r.closed.Store(true)
	if r.unregNetMon != nil {
		r.unregNetMon()
	}
	if err := r.downInterface(); err != nil {
		return err
	}
	if err := r.delIPRules(); err != nil {
		return err
	}
	if err := r.setNetfilterMode(netfilterOff); err != nil {
		return err
	}
	if err := r.delRoutes(); err != nil {
		return err
	}

	r.addrs = nil
	r.routes = nil
	r.localRoutes = nil

	return nil
}

// setupNetfilter initializes the NetfilterRunner in r.nfr. It expects r.nfr
// to be nil, or the current netfilter to be set to netfilterOff.
// kind should be either a linuxfw.FirewallMode, or the empty string for auto.
func (r *linuxRouter) setupNetfilter(kind string) error {
	r.netfilterKind = kind

	var err error
	r.nfr, err = linuxfw.New(r.logf, r.netfilterKind)
	if err != nil {
		return fmt.Errorf("could not create new netfilter: %w", err)
	}

	return nil
}

// Set implements the Router interface.
func (r *linuxRouter) Set(cfg *Config) error {
	var errs []error
	if cfg == nil {
		cfg = &shutdownConfig
	}

	if cfg.NetfilterKind != r.netfilterKind {
		if err := r.setNetfilterMode(netfilterOff); err != nil {
			err = fmt.Errorf("could not disable existing netfilter: %w", err)
			errs = append(errs, err)
		} else {
			r.nfr = nil
			if err := r.setupNetfilter(cfg.NetfilterKind); err != nil {
				errs = append(errs, err)
			}
		}
	}

	if err := r.setNetfilterMode(cfg.NetfilterMode); err != nil {
		errs = append(errs, err)
	}

	newLocalRoutes, err := cidrDiff("localRoute", r.localRoutes, cfg.LocalRoutes, r.addThrowRoute, r.delThrowRoute, r.logf)
	if err != nil {
		errs = append(errs, err)
	}
	r.localRoutes = newLocalRoutes

	newRoutes, err := cidrDiff("route", r.routes, cfg.Routes, r.addRoute, r.delRoute, r.logf)
	if err != nil {
		errs = append(errs, err)
	}
	r.routes = newRoutes

	newAddrs, err := cidrDiff("addr", r.addrs, cfg.LocalAddrs, r.addAddress, r.delAddress, r.logf)
	if err != nil {
		errs = append(errs, err)
	}
	r.addrs = newAddrs

	switch {
	case cfg.SNATSubnetRoutes == r.snatSubnetRoutes:
		// state already correct, nothing to do.
	case cfg.SNATSubnetRoutes:
		if err := r.addSNATRule(); err != nil {
			errs = append(errs, err)
		}
	default:
		if err := r.delSNATRule(); err != nil {
			errs = append(errs, err)
		}
	}
	r.snatSubnetRoutes = cfg.SNATSubnetRoutes

	// Issue 11405: enable IP forwarding on gokrazy.
	advertisingRoutes := len(cfg.SubnetRoutes) > 0
	if distro.Get() == distro.Gokrazy && advertisingRoutes {
		r.enableIPForwarding()
	}

	return multierr.New(errs...)
}

// UpdateMagicsockPort implements the Router interface.
func (r *linuxRouter) UpdateMagicsockPort(port uint16, network string) error {
	if r.nfr == nil {
		if err := r.setupNetfilter(r.netfilterKind); err != nil {
			return fmt.Errorf("could not setup netfilter: %w", err)
		}
	}

	var magicsockPort *uint16
	switch network {
	case "udp4":
		magicsockPort = &r.magicsockPortV4
	case "udp6":
		// Skip setting up MagicSock port if the host does not support
		// IPv6. MagicSock IPv6 port needs a filter rule to function. In
		// some cases (hosts with partial iptables support) filter
		// tables are not supported, so skip setting up the port for
		// those hosts too.
		if !r.getV6FilteringAvailable() {
			return nil
		}
		magicsockPort = &r.magicsockPortV6
	default:
		return fmt.Errorf("unsupported network %s", network)
	}

	// set the port, we'll make the firewall rule when netfilter turns back on
	if r.netfilterMode == netfilterOff {
		*magicsockPort = port
		return nil
	}

	if *magicsockPort == port {
		return nil
	}

	if *magicsockPort != 0 {
		if err := r.nfr.DelMagicsockPortRule(*magicsockPort, network); err != nil {
			return fmt.Errorf("del magicsock port rule: %w", err)
		}
	}

	if port != 0 {
		if err := r.nfr.AddMagicsockPortRule(*magicsockPort, network); err != nil {
			return fmt.Errorf("add magicsock port rule: %w", err)
		}
	}

	*magicsockPort = port
	return nil
}

// setNetfilterMode switches the router to the given netfilter
// mode. Netfilter state is created or deleted appropriately to
// reflect the new mode, and r.snatSubnetRoutes is updated to reflect
// the current state of subnet SNATing.
func (r *linuxRouter) setNetfilterMode(mode preftype.NetfilterMode) error {
	if distro.Get() == distro.Synology {
		mode = netfilterOff
	}

	if r.nfr == nil {
		var err error
		r.nfr, err = linuxfw.New(r.logf, r.netfilterKind)
		if err != nil {
			return err
		}
	}

	if r.netfilterMode == mode {
		return nil
	}

	// Depending on the netfilter mode we switch from and to, we may
	// have created the Tailscale netfilter chains. If so, we have to
	// go back through existing router state, and add the netfilter
	// rules for that state.
	//
	// This bool keeps track of whether the current state transition
	// is one that requires adding rules of existing state.
	reprocess := false

	switch mode {
	case netfilterOff:
		switch r.netfilterMode {
		case netfilterNoDivert:
			if err := r.nfr.DelBase(); err != nil {
				return err
			}
			if err := r.nfr.DelChains(); err != nil {
				r.logf("note: %v", err)
				// harmless, continue.
				// This can happen if someone left a ref to
				// this table somewhere else.
			}
		case netfilterOn:
			if err := r.nfr.DelHooks(r.logf); err != nil {
				return err
			}
			if err := r.nfr.DelBase(); err != nil {
				return err
			}
			if err := r.nfr.DelChains(); err != nil {
				r.logf("note: %v", err)
				// harmless, continue.
				// This can happen if someone left a ref to
				// this table somewhere else.
			}
		}
		r.snatSubnetRoutes = false
	case netfilterNoDivert:
		switch r.netfilterMode {
		case netfilterOff:
			reprocess = true
			if err := r.nfr.AddChains(); err != nil {
				return err
			}
			if err := r.nfr.AddBase(r.tunname); err != nil {
				return err
			}
			if r.magicsockPortV4 != 0 {
				if err := r.nfr.AddMagicsockPortRule(r.magicsockPortV4, "udp4"); err != nil {
					return fmt.Errorf("could not add magicsock port rule v4: %w", err)
				}
			}
			if r.magicsockPortV6 != 0 && r.getV6FilteringAvailable() {
				if err := r.nfr.AddMagicsockPortRule(r.magicsockPortV6, "udp6"); err != nil {
					return fmt.Errorf("could not add magicsock port rule v6: %w", err)
				}
			}
			r.snatSubnetRoutes = false
		case netfilterOn:
			if err := r.nfr.DelHooks(r.logf); err != nil {
				return err
			}
		}
	case netfilterOn:
		// Because of bugs in old version of iptables-compat,
		// we can't add a "-j ts-forward" rule to FORWARD
		// while ts-forward contains an "-m mark" rule. But
		// we can add the row *before* populating ts-forward.
		// So we have to delBase, then add the hooks,
		// then re-addBase, just in case.
		switch r.netfilterMode {
		case netfilterOff:
			reprocess = true
			if err := r.nfr.AddChains(); err != nil {
				return err
			}
			if err := r.nfr.DelBase(); err != nil {
				return err
			}
			// AddHooks adds the ts loopback rule.
			if err := r.nfr.AddHooks(); err != nil {
				return err
			}
			// AddBase adds base ts rules
			if err := r.nfr.AddBase(r.tunname); err != nil {
				return err
			}
			if r.magicsockPortV4 != 0 {
				if err := r.nfr.AddMagicsockPortRule(r.magicsockPortV4, "udp4"); err != nil {
					return fmt.Errorf("could not add magicsock port rule v4: %w", err)
				}
			}
			if r.magicsockPortV6 != 0 && r.getV6FilteringAvailable() {
				if err := r.nfr.AddMagicsockPortRule(r.magicsockPortV6, "udp6"); err != nil {
					return fmt.Errorf("could not add magicsock port rule v6: %w", err)
				}
			}
			r.snatSubnetRoutes = false
		case netfilterNoDivert:
			reprocess = true
			if err := r.nfr.DelBase(); err != nil {
				return err
			}
			if err := r.nfr.AddHooks(); err != nil {
				return err
			}
			if err := r.nfr.AddBase(r.tunname); err != nil {
				return err
			}
			r.snatSubnetRoutes = false
		}
	default:
		panic("unhandled netfilter mode")
	}

	r.netfilterMode = mode

	if !reprocess {
		return nil
	}

	for cidr := range r.addrs {
		if err := r.addLoopbackRule(cidr.Addr()); err != nil {
			return fmt.Errorf("error adding loopback rule: %w", err)
		}
	}

	return nil
}

// getV6FilteringAvailable returns true if the router is able to setup the
// required tailscale filter rules for IPv6.
func (r *linuxRouter) getV6FilteringAvailable() bool {
	return r.nfr.HasIPV6() && r.nfr.HasIPV6Filter()
}

// getV6Available returns true if the host supports IPv6.
func (r *linuxRouter) getV6Available() bool {
	return r.nfr.HasIPV6()
}

// addAddress adds an IP/mask to the tunnel interface. Fails if the
// address is already assigned to the interface, or if the addition
// fails.
func (r *linuxRouter) addAddress(addr netip.Prefix) error {
	if !r.getV6Available() && addr.Addr().Is6() {
		return nil
	}
	if r.useIPCommand() {
		if err := r.cmd.run("ip", "addr", "add", addr.String(), "dev", r.tunname); err != nil {
			return fmt.Errorf("adding address %q to tunnel interface: %w", addr, err)
		}
	} else {
		link, err := r.link()
		if err != nil {
			return fmt.Errorf("adding address %v, %w", addr, err)
		}
		if err := netlink.AddrReplace(link, nlAddrOfPrefix(addr)); err != nil {
			return fmt.Errorf("adding address %v from tunnel interface: %w", addr, err)
		}
	}
	if err := r.addLoopbackRule(addr.Addr()); err != nil {
		return err
	}
	return nil
}

// delAddress removes an IP/mask from the tunnel interface. Fails if
// the address is not assigned to the interface, or if the removal
// fails.
func (r *linuxRouter) delAddress(addr netip.Prefix) error {
	if !r.getV6Available() && addr.Addr().Is6() {
		return nil
	}
	if err := r.delLoopbackRule(addr.Addr()); err != nil {
		return err
	}
	if r.useIPCommand() {
		if err := r.cmd.run("ip", "addr", "del", addr.String(), "dev", r.tunname); err != nil {
			return fmt.Errorf("deleting address %q from tunnel interface: %w", addr, err)
		}
	} else {
		link, err := r.link()
		if err != nil {
			return fmt.Errorf("deleting address %v, %w", addr, err)
		}
		if err := netlink.AddrDel(link, nlAddrOfPrefix(addr)); err != nil {
			return fmt.Errorf("deleting address %v from tunnel interface: %w", addr, err)
		}
	}
	return nil
}

// addLoopbackRule adds a firewall rule to permit loopback traffic to
// a local Tailscale IP.
func (r *linuxRouter) addLoopbackRule(addr netip.Addr) error {
	if r.netfilterMode == netfilterOff {
		return nil
	}
	if addr.Is6() && !r.nfr.HasIPV6Filter() {
		return nil
	}

	if err := r.nfr.AddLoopbackRule(addr); err != nil {
		return err
	}
	return nil
}

// delLoopbackRule removes the firewall rule permitting loopback
// traffic to a Tailscale IP.
func (r *linuxRouter) delLoopbackRule(addr netip.Addr) error {
	if r.netfilterMode == netfilterOff {
		return nil
	}
	if addr.Is6() && !r.nfr.HasIPV6Filter() {
		return nil
	}

	if err := r.nfr.DelLoopbackRule(addr); err != nil {
		return err
	}
	return nil
}

// addRoute adds a route for cidr, pointing to the tunnel
// interface. Fails if the route already exists, or if adding the
// route fails.
func (r *linuxRouter) addRoute(cidr netip.Prefix) error {
	if !r.getV6Available() && cidr.Addr().Is6() {
		return nil
	}
	if r.useIPCommand() {
		return r.addRouteDef([]string{normalizeCIDR(cidr), "dev", r.tunname}, cidr)
	}
	linkIndex, err := r.linkIndex()
	if err != nil {
		return err
	}
	return netlink.RouteReplace(&netlink.Route{
		LinkIndex: linkIndex,
		Dst:       netipx.PrefixIPNet(cidr.Masked()),
		Table:     r.routeTable(),
	})
}

// addThrowRoute adds a throw route for the provided cidr.
// This has the effect that lookup in the routing table is terminated
// pretending that no route was found. Fails if the route already exists,
// or if adding the route fails.
func (r *linuxRouter) addThrowRoute(cidr netip.Prefix) error {
	if !r.ipRuleAvailable {
		return nil
	}
	if !r.getV6Available() && cidr.Addr().Is6() {
		return nil
	}
	if r.useIPCommand() {
		return r.addRouteDef([]string{"throw", normalizeCIDR(cidr)}, cidr)
	}
	err := netlink.RouteReplace(&netlink.Route{
		Dst:   netipx.PrefixIPNet(cidr.Masked()),
		Table: tailscaleRouteTable.Num,
		Type:  unix.RTN_THROW,
	})
	if err != nil {
		r.logf("THROW ERROR adding %v: %#v", cidr, err)
	}
	return err
}

func (r *linuxRouter) addRouteDef(routeDef []string, cidr netip.Prefix) error {
	if !r.getV6Available() && cidr.Addr().Is6() {
		return nil
	}
	args := append([]string{"ip", "route", "add"}, routeDef...)
	if r.ipRuleAvailable {
		args = append(args, "table", tailscaleRouteTable.ipCmdArg())
	}
	err := r.cmd.run(args...)
	if err == nil {
		return nil
	}

	// This is an ugly hack to detect failure to add a route that
	// already exists (as happens in when we're racing to add
	// kernel-maintained routes when enabling exit nodes w/o Local
	// LAN access, Issue 3060). Fortunately in the common case we
	// use netlink directly instead and don't exercise this code.
	if errCode(err) == 2 && strings.Contains(err.Error(), "RTNETLINK answers: File exists") {
		r.logf("ignoring route add of %v; already exists", cidr)
		return nil
	}
	return err
}

var (
	errESRCH  error = syscall.ESRCH
	errENOENT error = syscall.ENOENT
	errEEXIST error = syscall.EEXIST
)

// 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 netip.Prefix) error {
	if !r.getV6Available() && cidr.Addr().Is6() {
		return nil
	}
	if r.useIPCommand() {
		return r.delRouteDef([]string{normalizeCIDR(cidr), "dev", r.tunname}, cidr)
	}
	linkIndex, err := r.linkIndex()
	if err != nil {
		return err
	}
	err = netlink.RouteDel(&netlink.Route{
		LinkIndex: linkIndex,
		Dst:       netipx.PrefixIPNet(cidr.Masked()),
		Table:     r.routeTable(),
	})
	if errors.Is(err, errESRCH) {
		// Didn't exist to begin with.
		return nil
	}
	return err
}

// delThrowRoute removes the throw route for the cidr. Fails if the route
// doesn't exist, or if removing the route fails.
func (r *linuxRouter) delThrowRoute(cidr netip.Prefix) error {
	if !r.ipRuleAvailable {
		return nil
	}
	if !r.getV6Available() && cidr.Addr().Is6() {
		return nil
	}
	if r.useIPCommand() {
		return r.delRouteDef([]string{"throw", normalizeCIDR(cidr)}, cidr)
	}
	err := netlink.RouteDel(&netlink.Route{
		Dst:   netipx.PrefixIPNet(cidr.Masked()),
		Table: r.routeTable(),
		Type:  unix.RTN_THROW,
	})
	if errors.Is(err, errESRCH) {
		// Didn't exist to begin with.
		return nil
	}
	return err
}

func (r *linuxRouter) delRouteDef(routeDef []string, cidr netip.Prefix) error {
	if !r.getV6Available() && cidr.Addr().Is6() {
		return nil
	}
	args := append([]string{"ip", "route", "del"}, routeDef...)
	if r.ipRuleAvailable {
		args = append(args, "table", tailscaleRouteTable.ipCmdArg())
	}
	err := r.cmd.run(args...)
	if err != nil {
		ok, err := r.hasRoute(routeDef, cidr)
		if err != nil {
			r.logf("warning: error checking whether %v even exists after error deleting it: %v", err)
		} else {
			if !ok {
				r.logf("warning: tried to delete route %v but it was already gone; ignoring error", cidr)
				return nil
			}
		}
	}
	return err
}

func dashFam(ip netip.Addr) string {
	if ip.Is6() {
		return "-6"
	}
	return "-4"
}

func (r *linuxRouter) hasRoute(routeDef []string, cidr netip.Prefix) (bool, error) {
	args := append([]string{"ip", dashFam(cidr.Addr()), "route", "show"}, routeDef...)
	if r.ipRuleAvailable {
		args = append(args, "table", tailscaleRouteTable.ipCmdArg())
	}
	out, err := r.cmd.output(args...)
	if err != nil {
		return false, err
	}
	return len(out) > 0, nil
}

func (r *linuxRouter) link() (netlink.Link, error) {
	link, err := netlink.LinkByName(r.tunname)
	if err != nil {
		return nil, fmt.Errorf("failed to look up link %q: %w", r.tunname, err)
	}
	return link, nil
}

func (r *linuxRouter) linkIndex() (int, error) {
	// TODO(bradfitz): cache this? It doesn't change often, and on start-up
	// hundreds of addRoute calls to add /32s can happen quickly.
	link, err := r.link()
	if err != nil {
		return 0, err
	}
	return link.Attrs().Index, nil
}

// routeTable returns the route table to use.
func (r *linuxRouter) routeTable() int {
	if r.ipRuleAvailable {
		return tailscaleRouteTable.Num
	}
	return 0
}

// upInterface brings up the tunnel interface.
func (r *linuxRouter) upInterface() error {
	if r.useIPCommand() {
		return r.cmd.run("ip", "link", "set", "dev", r.tunname, "up")
	}
	link, err := r.link()
	if err != nil {
		return fmt.Errorf("bringing interface up, %w", err)
	}
	return netlink.LinkSetUp(link)
}

func (r *linuxRouter) enableIPForwarding() {
	sysctls := map[string]string{
		"net.ipv4.ip_forward":          "1",
		"net.ipv6.conf.all.forwarding": "1",
	}
	for k, v := range sysctls {
		if err := writeSysctl(k, v); err != nil {
			r.logf("warning: %v", k, v, err)
			continue
		}
		r.logf("sysctl(%v=%v): ok", k, v)
	}
}

func writeSysctl(key, val string) error {
	fn := "/proc/sys/" + strings.Replace(key, ".", "/", -1)
	if err := os.WriteFile(fn, []byte(val), 0644); err != nil {
		return fmt.Errorf("sysctl(%v=%v): %v", key, val, err)
	}
	return nil
}

// downInterface sets the tunnel interface administratively down.
func (r *linuxRouter) downInterface() error {
	if r.useIPCommand() {
		return r.cmd.run("ip", "link", "set", "dev", r.tunname, "down")
	}
	link, err := r.link()
	if err != nil {
		return fmt.Errorf("bringing interface down, %w", err)
	}
	return netlink.LinkSetDown(link)
}

// fixupWSLMTU sets the MTU on the eth0 interface to 1360 bytes if running under
// WSL, eth0 is the default route, and has the MTU 1280 bytes.
func (r *linuxRouter) fixupWSLMTU() {
	if !distro.IsWSL() {
		return
	}

	if r.useIPCommand() {
		r.logf("fixupWSLMTU: not implemented by ip command")
		return
	}

	link, err := netlink.LinkByName("eth0")
	if err != nil {
		r.logf("warning: fixupWSLMTU: could not open eth0: %v", err)
		return
	}

	routes, err := netlink.RouteGet(net.IPv4(8, 8, 8, 8))
	if err != nil || len(routes) == 0 {
		if err == nil {
			err = fmt.Errorf("none found")
		}
		r.logf("fixupWSLMTU: could not get default route: %v", err)
		return
	}

	if routes[0].LinkIndex != link.Attrs().Index {
		r.logf("fixupWSLMTU: default route is not via eth0")
		return
	}

	if link.Attrs().MTU == 1280 {
		if err := netlink.LinkSetMTU(link, 1360); err != nil {
			r.logf("warning: fixupWSLMTU: could not raise eth0 MTU: %v", err)
		}
	}
}

// addrFamily is an address family: IPv4 or IPv6.
type addrFamily byte

const (
	v4 = addrFamily(4)
	v6 = addrFamily(6)
)

func (f addrFamily) dashArg() string {
	switch f {
	case 4:
		return "-4"
	case 6:
		return "-6"
	}
	panic("illegal")
}

func (f addrFamily) netlinkInt() int {
	switch f {
	case 4:
		return netlink.FAMILY_V4
	case 6:
		return netlink.FAMILY_V6
	}
	panic("illegal")
}

func (r *linuxRouter) addrFamilies() []addrFamily {
	if r.getV6Available() {
		return []addrFamily{v4, v6}
	}
	return []addrFamily{v4}
}

// 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) 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.delIPRules(); err != nil {
		return err
	}

	return r.justAddIPRules()
}

// RouteTable is a Linux routing table: both its name and number.
// See /etc/iproute2/rt_tables.
type RouteTable struct {
	Name string
	Num  int
}

var routeTableByNumber = map[int]RouteTable{}

// IpCmdArg returns the string form of the table to pass to the "ip" command.
func (rt RouteTable) ipCmdArg() string {
	if rt.Num >= 253 {
		return rt.Name
	}
	return strconv.Itoa(rt.Num)
}

func newRouteTable(name string, num int) RouteTable {
	rt := RouteTable{name, num}
	routeTableByNumber[num] = rt
	return rt
}

// MustRouteTable returns the RouteTable with the given number key.
// It panics if the number is unknown because this result is a part
// of IP rule argument and we don't want to continue with an invalid
// argument with table no exist.
func mustRouteTable(num int) RouteTable {
	rt, ok := routeTableByNumber[num]
	if !ok {
		panic(fmt.Sprintf("unknown route table %v", num))
	}
	return rt
}

var (
	mainRouteTable    = newRouteTable("main", 254)
	defaultRouteTable = newRouteTable("default", 253)

	// tailscaleRouteTable is the routing table number for Tailscale
	// network routes. See addIPRules for the detailed policy routing
	// logic that ends up doing lookups within that table.
	//
	// NOTE(danderson): We chose 52 because those are the digits above the
	// letters "TS" on a qwerty keyboard, and 52 is sufficiently unlikely
	// to be picked by other software.
	//
	// NOTE(danderson): You might wonder why we didn't pick some
	// high table number like 5252, to further avoid the potential
	// for collisions with other software. Unfortunately,
	// Busybox's `ip` implementation believes that table numbers
	// are 8-bit integers, so for maximum compatibility we had to
	// stay in the 0-255 range even though linux itself supports
	// larger numbers. (but nowadays we use netlink directly and
	// aren't affected by the busybox binary's limitations)
	tailscaleRouteTable = newRouteTable("tailscale", 52)
)

// ipRules are the policy routing rules that Tailscale uses.
// The priority is the value represented here added to r.ipPolicyPrefBase,
// which is usually 5200.
//
// NOTE(apenwarr): We leave spaces between each pref number.
// This is so the sysadmin can override by inserting rules in
// between if they want.
//
// NOTE(apenwarr): This sequence seems complicated, right?
// If we could simply have a rule that said "match packets that
// *don't* have this fwmark", then we would only need to add one
// link to table 52 and we'd be done. Unfortunately, older kernels
// and 'ip rule' implementations (including busybox), don't support
// checking for the lack of a fwmark, only the presence. The technique
// below works even on very old kernels.
var ipRules = []netlink.Rule{
	// Packets from us, tagged with our fwmark, first try the kernel's
	// main routing table.
	{
		Priority: 10,
		Mark:     linuxfw.TailscaleBypassMarkNum,
		Table:    mainRouteTable.Num,
	},
	// ...and then we try the 'default' table, for correctness,
	// even though it's been empty on every Linux system I've ever seen.
	{
		Priority: 30,
		Mark:     linuxfw.TailscaleBypassMarkNum,
		Table:    defaultRouteTable.Num,
	},
	// If neither of those matched (no default route on this system?)
	// then packets from us should be aborted rather than falling through
	// to the tailscale routes, because that would create routing loops.
	{
		Priority: 50,
		Mark:     linuxfw.TailscaleBypassMarkNum,
		Type:     unix.RTN_UNREACHABLE,
	},
	// If we get to this point, capture all packets and send them
	// through to the tailscale route table. For apps other than us
	// (ie. with no fwmark set), this is the first routing table, so
	// it takes precedence over all the others, ie. VPN routes always
	// beat non-VPN routes.
	{
		Priority: 70,
		Table:    tailscaleRouteTable.Num,
	},
	// If that didn't match, then non-fwmark packets fall through to the
	// usual rules (pref 32766 and 32767, ie. main and default).
}

// justAddIPRules adds policy routing rule without deleting any first.
func (r *linuxRouter) justAddIPRules() error {
	if !r.ipRuleAvailable {
		return nil
	}
	if r.useIPCommand() {
		return r.addIPRulesWithIPCommand()
	}
	var errAcc error
	for _, family := range r.addrFamilies() {

		for _, ru := range ipRules {
			// Note: r is a value type here; safe to mutate it.
			ru.Family = family.netlinkInt()
			if ru.Mark != 0 {
				ru.Mask = linuxfw.TailscaleFwmarkMaskNum
			}
			ru.Goto = -1
			ru.SuppressIfgroup = -1
			ru.SuppressPrefixlen = -1
			ru.Flow = -1
			ru.Priority += r.ipPolicyPrefBase

			err := netlink.RuleAdd(&ru)
			if errors.Is(err, errEEXIST) {
				// Ignore dups.
				continue
			}
			if err != nil && errAcc == nil {
				errAcc = err
			}
		}
	}
	return errAcc
}

func (r *linuxRouter) addIPRulesWithIPCommand() error {
	rg := newRunGroup(nil, r.cmd)

	for _, family := range r.addrFamilies() {
		for _, rule := range ipRules {
			args := []string{
				"ip", family.dashArg(),
				"rule", "add",
				"pref", strconv.Itoa(rule.Priority + r.ipPolicyPrefBase),
			}
			if rule.Mark != 0 {
				if r.fwmaskWorks {
					args = append(args, "fwmark", fmt.Sprintf("0x%x/%s", rule.Mark, linuxfw.TailscaleFwmarkMask))
				} else {
					args = append(args, "fwmark", fmt.Sprintf("0x%x", rule.Mark))
				}
			}
			if rule.Table != 0 {
				args = append(args, "table", mustRouteTable(rule.Table).ipCmdArg())
			}
			if rule.Type == unix.RTN_UNREACHABLE {
				args = append(args, "type", "unreachable")
			}
			rg.Run(args...)
		}
	}

	return rg.ErrAcc
}

// delRoutes removes any local routes that we added that would not be
// cleaned up on interface down.
func (r *linuxRouter) delRoutes() error {
	for rt := range r.localRoutes {
		if err := r.delThrowRoute(rt); err != nil {
			r.logf("failed to delete throw route(%q): %v", rt, err)
		}
	}
	return nil
}

// delIPRules removes the policy routing rules that avoid
// tailscaled routing loops, if it exists.
func (r *linuxRouter) delIPRules() error {
	if !r.ipRuleAvailable {
		return nil
	}
	if r.useIPCommand() {
		return r.delIPRulesWithIPCommand()
	}
	var errAcc error
	for _, family := range r.addrFamilies() {
		for _, ru := range ipRules {
			// Note: r is a value type here; safe to mutate it.
			// When deleting rules, we want to be a bit specific (mention which
			// table we were routing to) but not *too* specific (fwmarks, etc).
			// That leaves us some flexibility to change these values in later
			// versions without having ongoing hacks for every possible
			// combination.
			ru.Family = family.netlinkInt()
			ru.Mark = -1
			ru.Mask = -1
			ru.Goto = -1
			ru.SuppressIfgroup = -1
			ru.SuppressPrefixlen = -1
			ru.Priority += r.ipPolicyPrefBase

			err := netlink.RuleDel(&ru)
			if errors.Is(err, errENOENT) {
				// Didn't exist to begin with.
				continue
			}
			if err != nil && errAcc == nil {
				errAcc = err
			}
		}
	}
	return errAcc
}

func (r *linuxRouter) delIPRulesWithIPCommand() error {
	// 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.
	//
	// Some older versions of iproute2 also return error code 254 for
	// unknown rules during deletion.
	rg := newRunGroup([]int{2, 254}, r.cmd)

	for _, family := range r.addrFamilies() {
		// When deleting rules, we want to be a bit specific (mention which
		// table we were routing to) but not *too* specific (fwmarks, etc).
		// That leaves us some flexibility to change these values in later
		// versions without having ongoing hacks for every possible
		// combination.
		for _, rule := range ipRules {
			args := []string{
				"ip", family.dashArg(),
				"rule", "del",
				"pref", strconv.Itoa(rule.Priority + r.ipPolicyPrefBase),
			}
			if rule.Table != 0 {
				args = append(args, "table", mustRouteTable(rule.Table).ipCmdArg())
			} else {
				args = append(args, "type", "unreachable")
			}
			rg.Run(args...)
		}
	}

	return rg.ErrAcc
}

// addSNATRule adds a netfilter rule to SNAT traffic destined for
// local subnets.
func (r *linuxRouter) addSNATRule() error {
	if r.netfilterMode == netfilterOff {
		return nil
	}

	if err := r.nfr.AddSNATRule(); err != nil {
		return err
	}
	return nil
}

// delSNATRule removes the netfilter rule to SNAT traffic destined for
// local subnets. Fails if the rule does not exist.
func (r *linuxRouter) delSNATRule() error {
	if r.netfilterMode == netfilterOff {
		return nil
	}

	if err := r.nfr.DelSNATRule(); err != nil {
		return err
	}
	return nil
}

// cidrDiff calls add and del as needed to make the set of prefixes in
// old and new match. Returns a map reflecting the actual new state
// (which may be somewhere in between old and new if some commands
// failed), and any error encountered while reconfiguring.
func cidrDiff(kind string, old map[netip.Prefix]bool, new []netip.Prefix, add, del func(netip.Prefix) error, logf logger.Logf) (map[netip.Prefix]bool, error) {
	newMap := make(map[netip.Prefix]bool, len(new))
	for _, cidr := range new {
		newMap[cidr] = true
	}

	// ret starts out as a copy of old, and updates as we
	// add/delete. That way we can always return it and have it be the
	// true state of what we've done so far.
	ret := make(map[netip.Prefix]bool, len(old))
	for cidr := range old {
		ret[cidr] = true
	}

	// We want to add before we delete, so that if there is no overlap, we don't
	// end up in a state where we have no addresses on an interface as that
	// results in other kernel entities (like routes) pointing to that interface
	// to also be deleted.
	var addFail []error
	for cidr := range newMap {
		if old[cidr] {
			continue
		}
		if err := add(cidr); err != nil {
			logf("%s add failed: %v", kind, err)
			addFail = append(addFail, err)
		} else {
			ret[cidr] = true
		}
	}

	if len(addFail) == 1 {
		return ret, addFail[0]
	}
	if len(addFail) > 0 {
		return ret, fmt.Errorf("%d add %s failures; first was: %w", len(addFail), kind, addFail[0])
	}

	var delFail []error
	for cidr := range old {
		if newMap[cidr] {
			continue
		}
		if err := del(cidr); err != nil {
			logf("%s del failed: %v", kind, err)
			delFail = append(delFail, err)
		} else {
			delete(ret, cidr)
		}
	}
	if len(delFail) == 1 {
		return ret, delFail[0]
	}
	if len(delFail) > 0 {
		return ret, fmt.Errorf("%d delete %s failures; first was: %w", len(delFail), kind, delFail[0])
	}

	return ret, nil
}

// normalizeCIDR returns cidr as an ip/mask string, with the host bits
// of the IP address zeroed out.
func normalizeCIDR(cidr netip.Prefix) string {
	return cidr.Masked().String()
}

// cleanup removes all the rules and routes that were added by the linux router.
// The function calls cleanup for both iptables and nftables since which ever
// netfilter runner is used, the cleanup function for the other one doesn't do anything.
func cleanup(logf logger.Logf, interfaceName string) {
	if interfaceName != "userspace-networking" {
		linuxfw.IPTablesCleanup(logf)
		linuxfw.NfTablesCleanUp(logf)
	}
}

// Checks if the running openWRT system is using mwan3, based on the heuristic
// of the config file being present as well as a policy rule with a specific
// priority (2000 + 1 - first interface mwan3 manages) and non-zero mark.
func checkOpenWRTUsingMWAN3() (bool, error) {
	if distro.Get() != distro.OpenWrt {
		return false, nil
	}

	if _, err := os.Stat("/etc/config/mwan3"); err != nil {
		if os.IsNotExist(err) {
			return false, nil
		}
		return false, err
	}

	rules, err := netlink.RuleList(netlink.FAMILY_V4)
	if err != nil {
		return false, err
	}
	for _, r := range rules {
		// We want to match on a rule like this:
		//    2001:	from all fwmark 0x100/0x3f00 lookup 1
		//
		// We dont match on the mask because it can vary, or the
		// table because I'm not sure if it can vary.
		if r.Priority >= 2001 && r.Priority <= 2004 && r.Mark != 0 {
			return true, nil
		}
	}

	return false, nil
}

func nlAddrOfPrefix(p netip.Prefix) *netlink.Addr {
	return &netlink.Addr{
		IPNet: netipx.PrefixIPNet(p),
	}
}