race when updating labels and store topology runs concurrently

[you06]

Bug Report

Please answer these questions before submitting your issue. Thanks!

Found by @lcwangchao 's great code review.

1. Minimal reproduce step (Required)

1. Add a test into pkg/server/handler/tests/http_handler_test.go.

func TestSetLabelsConcurrentWithStoreTopology(t *testing.T) {
	ts := createBasicHTTPHandlerTestSuite()
	ctx, cancel := context.WithCancel(context.Background())
	defer cancel()

	ts.startServer(t)
	defer ts.stopServer(t)

	integration.BeforeTestExternal(t)
	cluster := integration.NewClusterV3(t, &integration.ClusterConfig{Size: 1})
	defer cluster.Terminate(t)
	client := cluster.RandClient()
	infosync.SetEtcdClient(client)

	ts.domain.InfoSyncer().Restart(ctx)
	ts.domain.InfoSyncer().RestartTopology(ctx)

	testUpdateLabels := func() {
		labels := map[string]string{}
		labels["zone"] = fmt.Sprintf("z-%v", rand.Intn(100000))
		buffer := bytes.NewBuffer([]byte{})
		require.Nil(t, json.NewEncoder(buffer).Encode(labels))
		resp, err := ts.PostStatus("/labels", "application/json", buffer)
		require.NoError(t, err)
		require.NotNil(t, resp)
		defer func() {
			require.NoError(t, resp.Body.Close())
		}()
		require.Equal(t, http.StatusOK, resp.StatusCode)
		newLabels := config.GetGlobalConfig().Labels
		require.Equal(t, newLabels, labels)
	}
	testStoreTopology := func() {
		require.NoError(t, ts.domain.InfoSyncer().StoreTopologyInfo(context.Background()))
	}

	done := make(chan struct{})
	go func() {
		for {
			select {
			case <-done:
				return
			default:
				testStoreTopology()
			}
		}
	}()
	for i := 0; i < 100; i++ {
		testUpdateLabels()
	}
	close(done)

	// reset the global variable
	config.UpdateGlobal(func(conf *config.Config) {
		conf.Labels = map[string]string{}
	})
}

2. Run the test with race detection.

go test ./pkg/server/handler/tests -run TestSetLabelsConcurrentWithStoreTopology -race

2. What did you expect to see? (Required)

Test pass.

3. What did you see instead (Required)

WARNING: DATA RACE
Write at 0x00c00a78ece8 by goroutine 1975:
  github.com/pingcap/tidb/pkg/domain/infosync.UpdateServerLabel()
      github.com/pingcap/tidb/pkg/domain/infosync/info.go:443 +0x33b
  github.com/pingcap/tidb/pkg/server/handler/tikvhandler.LabelHandler.ServeHTTP()
      github.com/pingcap/tidb/pkg/server/handler/tikvhandler/tikv_handler.go:2011 +0x55a
  github.com/pingcap/tidb/pkg/server/handler/tikvhandler.(*LabelHandler).ServeHTTP()
      <autogenerated>:1 +0x58
  github.com/gorilla/mux.(*Router).ServeHTTP()
      github.com/gorilla/mux@v1.8.1/mux.go:212 +0x371
  net/http.(*ServeMux).ServeHTTP()
      net/http/server.go:2747 +0x255
  github.com/pingcap/tidb/pkg/server/internal/util.CorsHandler.ServeHTTP()
      github.com/pingcap/tidb/pkg/server/internal/util/util.go:219 +0x29e
  github.com/pingcap/tidb/pkg/server/internal/util.(*CorsHandler).ServeHTTP()
      <autogenerated>:1 +0x74
  net/http.serverHandler.ServeHTTP()
      net/http/server.go:3210 +0x2a1
  net/http.(*conn).serve()
      net/http/server.go:2092 +0x12a4
  net/http.(*Server).Serve.gowrap3()
      net/http/server.go:3360 +0x4f

Previous read at 0x00c00a78ece8 by goroutine 1928:
  reflect.typedmemmove()
      runtime/mbarrier.go:225 +0x0
  reflect.copyVal()
      reflect/value.go:2048 +0x5b
  reflect.(*MapIter).Value()
      reflect/value.go:1945 +0x124
  encoding/json.mapEncoder.encode()
      encoding/json/encode.go:760 +0x61a
  encoding/json.mapEncoder.encode-fm()
      <autogenerated>:1 +0x84
  encoding/json.structEncoder.encode()
      encoding/json/encode.go:715 +0x2bd
  encoding/json.structEncoder.encode-fm()
      <autogenerated>:1 +0xe4
  encoding/json.(*encodeState).reflectValue()
      encoding/json/encode.go:322 +0x83
  encoding/json.(*encodeState).marshal()
      encoding/json/encode.go:298 +0xea
  encoding/json.Marshal()
      encoding/json/encode.go:164 +0x12b
  github.com/pingcap/tidb/pkg/domain/infosync.(*InfoSyncer).StoreTopologyInfo()
      github.com/pingcap/tidb/pkg/domain/infosync/info.go:724 +0xf5
  github.com/pingcap/tidb/pkg/server/handler/tests.TestSetLabelsConcurrentWithStoreTopology.func2()
      github.com/pingcap/tidb/pkg/server/handler/tests/http_handler_test.go:1564 +0x84
  github.com/pingcap/tidb/pkg/server/handler/tests.TestSetLabelsConcurrentWithStoreTopology.func3()
      github.com/pingcap/tidb/pkg/server/handler/tests/http_handler_test.go:1574 +0x35

4. What is your TiDB version? (Required)

nightly

[you06]

BTW, I'm using SetEtcdClient in test, and there is also a data race in this function(L1333 and L810). We'd better deprecate SetEtcdClient, and replace it with a failpoint injection.

tidb/pkg/domain/infosync/info.go

Lines 1327 to 1334 in 1fa2955

	func SetEtcdClient(etcdCli *clientv3.Client) {
	is, err := getGlobalInfoSyncer()
	if err != nil {
	return
	}
	is.etcdCli = etcdCli
	}

tidb/pkg/domain/infosync/info.go

Lines 828 to 834 in 1fa2955

	// Done returns a channel that closes when the info syncer is no longer being refreshed.
	func (is *InfoSyncer) Done() <-chan struct{} {
	if is.etcdCli == nil {
	return make(chan struct{}, 1)
	}
	return is.session.Done()
	}

[you06]

I prefer moving the Labels out of the ServerInfo and keep it immutable and static as the comment states.

tidb/pkg/domain/infosync/info.go

Line 133 in 842b66e

// It will not be updated when tidb-server running. So please only put static information in ServerInfo struct.