net/dns: fix infinite loop when run on Amazon Linux 2023

This fixes an infinite loop caused by the configuration of
systemd-resolved on Amazon Linux 2023 and how that interacts with
Tailscale's "direct" mode. We now drop the Tailscale service IP from the
OS's "base configuration" when we detect this configuration.

Updates #7816

Signed-off-by: Andrew Dunham <andrew@du.nham.ca>
Change-Id: I73a4ea8e65571eb368c7e179f36af2c049a588ee

net/dns/direct.go

@@ -16,11 +16,13 @@ import (
 	"os/exec"
 	"path/filepath"
 	"runtime"
+	"slices"
 	"strings"
 	"sync"
 	"time"
 
 	"tailscale.com/net/dns/resolvconffile"
+	"tailscale.com/net/tsaddr"
 	"tailscale.com/types/logger"
 	"tailscale.com/util/dnsname"
 	"tailscale.com/version/distro"
@@ -371,7 +373,33 @@ func (m *directManager) GetBaseConfig() (OSConfig, error) {
 		fileToRead = backupConf
 	}
 
-	return m.readResolvFile(fileToRead)
+	oscfg, err := m.readResolvFile(fileToRead)
+	if err != nil {
+		return OSConfig{}, err
+	}
+
+	// On some systems, the backup configuration file is actually a
+	// symbolic link to something owned by another DNS service (commonly,
+	// resolved). Thus, it can be updated out from underneath us to contain
+	// the Tailscale service IP, which results in an infinite loop of us
+	// trying to send traffic to resolved, which sends back to us, and so
+	// on. To solve this, drop the Tailscale service IP from the base
+	// configuration; we do this in all situations since there's
+	// essentially no world where we want to forward to ourselves.
+	//
+	// See: https://github.com/tailscale/tailscale/issues/7816
+	var removed bool
+	oscfg.Nameservers = slices.DeleteFunc(oscfg.Nameservers, func(ip netip.Addr) bool {
+		if ip == tsaddr.TailscaleServiceIP() || ip == tsaddr.TailscaleServiceIPv6() {
+			removed = true
+			return true
+		}
+		return false
+	})
+	if removed {
+		m.logf("[v1] dropped Tailscale IP from base config that was a symlink")
+	}
+	return oscfg, nil
 }
 
 func (m *directManager) Close() error {