[client] Remove the empty GPO DNS policy store on Windows teardown (#7563)

2026-09-22 14:49:10 +02:00 · 2026-09-22 12:44:10 +02:00
parent ee2344502e
commit f6109a3395
2 changed files with 213 additions and 14 deletions
@@ -124,19 +124,9 @@ func newHostManager(wgInterface WGIface) (*registryConfigurator, error) {
 		return nil, err
 	}

-	var useGPO bool
-	k, err := registry.OpenKey(registry.LOCAL_MACHINE, GPODNSPolicyConfigRoot, registry.QUERY_VALUE)
-	if err != nil {
-		log.Debugf("failed to open GPO DNS policy root: %v", err)
-	} else {
-		closer(k)
-		useGPO = true
-		log.Infof("detected GPO DNS policy configuration, using policy store")
-	}
-
 	configurator := &registryConfigurator{
 		guid: guid,
-		gpo:  useGPO,
+		gpo:  useGPOPolicyStore(),
 	}

 	origNameservers, err := configurator.captureOriginalNameservers()
@@ -576,14 +566,22 @@ func (r *registryConfigurator) setInterfaceRegistryKeyStringValue(key, value str
 	return nil
 }

+// deleteInterfaceRegistryKeyProperty removes a value from the interface key.
+// A value that is already gone, or an interface key that is, is not an error:
+// the caller asked for the value not to be there, and a cleanup that runs twice
+// has to reach its later steps on the second run as well.
 func (r *registryConfigurator) deleteInterfaceRegistryKeyProperty(propertyKey string) error {
 	regKey, err := r.getInterfaceRegistryKey()
-	if err != nil {
+	switch {
+	case errors.Is(err, registry.ErrNotExist), errors.Is(err, syscall.ERROR_PATH_NOT_FOUND):
+		log.Debugf("interface key of %s does not exist, nothing to delete %s from", r.guid, propertyKey)
+		return nil
+	case err != nil:
 		return fmt.Errorf("get interface registry key: %w", err)
 	}
 	defer closer(regKey)

-	if err := regKey.DeleteValue(propertyKey); err != nil {
+	if err := regKey.DeleteValue(propertyKey); err != nil && !errors.Is(err, registry.ErrNotExist) {
 		return fmt.Errorf("delete registry key %s: %w", propertyKey, err)
 	}
 	return nil
@@ -612,7 +610,12 @@ func (r *registryConfigurator) restoreHostDNS() error {

 	go r.flushDNSCache()

-	return nil
+	// Last, and only on the way out, once no rule of ours is left: during a
+	// session the store is where the rules of this run live, and emptying it
+	// mid-session would have the next rule recreate it anyway. Propagated so a
+	// failure keeps the shutdown state for the next run to retry, rather than
+	// leaving the store to hold up every rule change from here on.
+	return removeEmptyGPOPolicyStore()
 }

 // removeDNSMatchPolicies deletes every NRPT rule this client may have created,
@@ -651,6 +654,73 @@ func (r *registryConfigurator) restoreUncleanShutdownDNS() error {
 	return r.restoreHostDNS()
 }

+// useGPOPolicyStore reports whether NRPT rules have to go into the group policy
+// store, and clears an empty one out of the way first.
+//
+// The order is the point. A store left empty by an earlier run would otherwise
+// decide this run too, sending its rules somewhere the resolver only reads when
+// the policy engine next applies DNS client policy. Removing it before the
+// choice is made leaves the local store authoritative for the whole session,
+// including the first one after an upgrade.
+func useGPOPolicyStore() bool {
+	if err := removeEmptyGPOPolicyStore(); err != nil {
+		// Nothing to retry against here: the worst case is the run going
+		// through the group policy store, which is where it would have gone
+		// before this check existed.
+		log.Warnf("%v", err)
+	}
+
+	k, err := registry.OpenKey(registry.LOCAL_MACHINE, GPODNSPolicyConfigRoot, registry.QUERY_VALUE)
+	if err != nil {
+		log.Debugf("failed to open GPO DNS policy root: %v", err)
+		return false
+	}
+	closer(k)
+
+	log.Infof("detected GPO DNS policy configuration, using policy store")
+	return true
+}
+
+// removeEmptyGPOPolicyStore deletes the group policy DnsPolicyConfig key once
+// nothing is left in it. The key survives the deletion of the last rule it
+// held, and the client treats its presence as "group policy configures the
+// NRPT", so an empty one left behind keeps every later run writing rules there.
+// Rules in that store reach the resolver only when the policy engine next
+// applies DNS client policy, and a rule this client writes belongs to no GPO,
+// so nothing schedules that application: both adding and removing a rule are
+// held up by a minute or more, and for a removal that is a catch-all rule
+// resolving every name over an interface that no longer exists. With the store
+// absent the local one is authoritative and a change applies at once.
+//
+// A store that still holds rules, values or subkeys of somebody else's is left
+// alone.
+func removeEmptyGPOPolicyStore() error {
+	k, err := registry.OpenKey(registry.LOCAL_MACHINE, GPODNSPolicyConfigRoot, registry.QUERY_VALUE)
+	switch {
+	case errors.Is(err, registry.ErrNotExist), errors.Is(err, syscall.ERROR_PATH_NOT_FOUND):
+		return nil
+	case err != nil:
+		return fmt.Errorf("open HKEY_LOCAL_MACHINE\\%s: %w", GPODNSPolicyConfigRoot, err)
+	}
+
+	info, err := k.Stat()
+	closer(k)
+	if err != nil {
+		return fmt.Errorf("stat HKEY_LOCAL_MACHINE\\%s: %w", GPODNSPolicyConfigRoot, err)
+	}
+
+	if info.SubKeyCount != 0 || info.ValueCount != 0 {
+		return nil
+	}
+
+	if err := registry.DeleteKey(registry.LOCAL_MACHINE, GPODNSPolicyConfigRoot); err != nil {
+		return fmt.Errorf("delete empty HKEY_LOCAL_MACHINE\\%s: %w", GPODNSPolicyConfigRoot, err)
+	}
+
+	log.Infof("removed the empty GPO DNS policy store, leaving the local one authoritative")
+	return nil
+}
+
 // listNRPTRuleKeys returns the names of our NRPT rule keys under a policy store
 // root. An absent root holds nothing to clean up, which is the normal state of
 // the GPO store on a machine without DNS Client policy.
@@ -8,6 +8,8 @@ import (
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 	"golang.org/x/sys/windows/registry"
+
+	"github.com/netbirdio/netbird/client/internal/winregistry"
 )

 // TestNRPTEntriesCleanupOnConfigChange tests that old NRPT entries are properly cleaned up
@@ -405,3 +407,130 @@ func TestNRPTDomainBatching(t *testing.T) {
 		})
 	}
 }
+
+// TestRemoveEmptyGPOPolicyStore verifies that cleanup takes the GPO policy
+// store itself with it once our rules are gone, since the store existing keeps
+// the local one from being applied, and that a store with somebody else's rule
+// in it is left alone.
+func TestRemoveEmptyGPOPolicyStore(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping registry integration test in short mode")
+	}
+
+	t.Cleanup(func() { cleanupRegistryKeys(t) })
+	cleanupRegistryKeys(t)
+
+	testIP := netip.MustParseAddr("100.64.0.1")
+	cfg := &registryConfigurator{gpo: true}
+
+	// a store holding a rule of ours is kept, because the rule is still applied
+	require.NoError(t, cfg.addDNSMatchPolicy([]string{".example.com"}, testIP))
+	exists, err := registryKeyExists(gpoDnsPolicyConfigMatchPath + "-0")
+	require.NoError(t, err)
+	require.True(t, exists, "Should write the rule to the GPO policy store")
+
+	require.NoError(t, removeEmptyGPOPolicyStore())
+	exists, err = registryKeyExists(GPODNSPolicyConfigRoot)
+	require.NoError(t, err)
+	assert.True(t, exists, "Should keep a policy store that still holds a rule")
+
+	// once the rules are gone the store goes with them
+	require.NoError(t, cfg.removeDNSMatchPolicies())
+	require.NoError(t, removeEmptyGPOPolicyStore())
+
+	exists, err = registryKeyExists(GPODNSPolicyConfigRoot)
+	require.NoError(t, err)
+	assert.False(t, exists, "Should remove the GPO policy store once it is empty")
+
+	// A store is not ours to remove while somebody else has a rule in it. The
+	// rule is written volatile like our own: the rules above created the parent
+	// chain volatile, and Windows refuses a stable subkey under a volatile
+	// parent.
+	foreignRule := GPODNSPolicyConfigRoot + `\{2A3B4C5D-6E7F-4041-8283-84858687888A}`
+	foreignKey, _, err := winregistry.CreateVolatileKey(registry.LOCAL_MACHINE, foreignRule, registry.SET_VALUE)
+	require.NoError(t, err, "Should create a foreign GPO rule")
+	foreignKey.Close()
+	t.Cleanup(func() {
+		_ = registry.DeleteKey(registry.LOCAL_MACHINE, foreignRule)
+		_ = registry.DeleteKey(registry.LOCAL_MACHINE, GPODNSPolicyConfigRoot)
+	})
+
+	require.NoError(t, cfg.removeDNSMatchPolicies())
+	require.NoError(t, removeEmptyGPOPolicyStore())
+
+	exists, err = registryKeyExists(foreignRule)
+	require.NoError(t, err)
+	assert.True(t, exists, "Should not remove a foreign rule")
+	exists, err = registryKeyExists(GPODNSPolicyConfigRoot)
+	require.NoError(t, err)
+	assert.True(t, exists, "Should keep a policy store that still holds a foreign rule")
+}
+
+// TestDeleteInterfaceRegistryKeyPropertyTwice verifies that removing a value
+// that is already gone, or one on an interface key that is, reports success.
+// Teardown runs again after a failed cleanup, and the steps that follow this
+// one have to be reached on that second run.
+func TestDeleteInterfaceRegistryKeyPropertyTwice(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping registry integration test in short mode")
+	}
+
+	testGUID := "{12345678-1234-1234-1234-123456789ABC}"
+	interfacePath := InterfaceConfigPath + `\` + testGUID
+	testKey, _, err := registry.CreateKey(registry.LOCAL_MACHINE, interfacePath, registry.SET_VALUE)
+	require.NoError(t, err, "Should create test interface registry key")
+	testKey.Close()
+	t.Cleanup(func() {
+		_ = registry.DeleteKey(registry.LOCAL_MACHINE, interfacePath)
+	})
+
+	cfg := &registryConfigurator{guid: testGUID}
+
+	require.NoError(t, cfg.setInterfaceRegistryKeyStringValue(interfaceConfigSearchListKey, "example.com"))
+	require.NoError(t, cfg.deleteInterfaceRegistryKeyProperty(interfaceConfigSearchListKey))
+	assert.NoError(t, cfg.deleteInterfaceRegistryKeyProperty(interfaceConfigSearchListKey),
+		"Should report success for a value that is already gone")
+
+	// and with the interface key itself gone, as it is once the adapter is
+	require.NoError(t, registry.DeleteKey(registry.LOCAL_MACHINE, interfacePath))
+	assert.NoError(t, cfg.deleteInterfaceRegistryKeyProperty(interfaceConfigSearchListKey),
+		"Should report success when the interface key does not exist")
+}
+
+// TestUseGPOPolicyStoreClearsEmptyStore verifies that the store is cleared
+// before it is consulted, so an empty one left by an earlier run does not send
+// this run's rules to the group policy store. A store somebody else has a rule
+// in still decides where the rules go.
+func TestUseGPOPolicyStoreClearsEmptyStore(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping registry integration test in short mode")
+	}
+
+	t.Cleanup(func() { cleanupRegistryKeys(t) })
+	cleanupRegistryKeys(t)
+
+	// the leftover an earlier run used to keep, which the client read as
+	// "group policy configures the NRPT" for every run after it
+	emptyStore, _, err := winregistry.CreateVolatileKey(registry.LOCAL_MACHINE, GPODNSPolicyConfigRoot, registry.SET_VALUE)
+	require.NoError(t, err, "Should create the GPO policy store")
+	emptyStore.Close()
+
+	assert.False(t, useGPOPolicyStore(), "An empty store should not decide where the rules go")
+	exists, err := registryKeyExists(GPODNSPolicyConfigRoot)
+	require.NoError(t, err)
+	assert.False(t, exists, "Should clear the empty store before consulting it")
+
+	foreignRule := GPODNSPolicyConfigRoot + `\{2A3B4C5D-6E7F-4041-8283-84858687888A}`
+	foreignKey, _, err := winregistry.CreateVolatileKey(registry.LOCAL_MACHINE, foreignRule, registry.SET_VALUE)
+	require.NoError(t, err, "Should create a foreign GPO rule")
+	foreignKey.Close()
+	t.Cleanup(func() {
+		_ = registry.DeleteKey(registry.LOCAL_MACHINE, foreignRule)
+		_ = registry.DeleteKey(registry.LOCAL_MACHINE, GPODNSPolicyConfigRoot)
+	})
+
+	assert.True(t, useGPOPolicyStore(), "A store holding a rule should decide where the rules go")
+	exists, err = registryKeyExists(GPODNSPolicyConfigRoot)
+	require.NoError(t, err)
+	assert.True(t, exists, "Should keep a store that holds a rule")
+}