br/operator: fix adapt env for snapshot backup stuck when encountered error (pingcap#52607) (pingcap#56717)

close pingcap#52049

Author: ti-chi-bot

br/pkg/backup/prepare_snap/BUILD.bazel

@@ -17,6 +17,7 @@ go_library(
         "@com_github_docker_go_units//:go-units",
         "@com_github_google_btree//:btree",
         "@com_github_pingcap_errors//:errors",
+        "@com_github_pingcap_failpoint//:failpoint",
         "@com_github_pingcap_kvproto//pkg/brpb",
         "@com_github_pingcap_kvproto//pkg/errorpb",
         "@com_github_pingcap_kvproto//pkg/metapb",

br/pkg/backup/prepare_snap/prepare.go

@@ -22,6 +22,7 @@ import (
 
 	"github.com/google/btree"
 	"github.com/pingcap/errors"
+	"github.com/pingcap/failpoint"
 	brpb "github.com/pingcap/kvproto/pkg/brpb"
 	"github.com/pingcap/kvproto/pkg/metapb"
 	"github.com/pingcap/log"
@@ -453,6 +454,9 @@ func (p *Preparer) pushWaitApply(reqs pendingRequests, region Region) {
 // PrepareConnections prepares the connections for each store.
 // This will pause the admin commands for each store.
 func (p *Preparer) PrepareConnections(ctx context.Context) error {
+	failpoint.Inject("PrepareConnectionsErr", func() {
+		failpoint.Return(errors.New("mock PrepareConnectionsErr"))
+	})
 	log.Info("Preparing connections to stores.")
 	stores, err := p.env.GetAllLiveStores(ctx)
 	if err != nil {

br/pkg/task/operator/BUILD.bazel

@@ -16,6 +16,7 @@ go_library(
         "//br/pkg/task",
         "//br/pkg/utils",
         "@com_github_pingcap_errors//:errors",
+        "@com_github_pingcap_failpoint//:failpoint",
         "@com_github_pingcap_log//:log",
         "@com_github_spf13_pflag//:pflag",
         "@com_github_tikv_client_go_v2//tikv",

br/pkg/task/operator/cmd.go

@@ -5,14 +5,12 @@ package operator
 import (
 	"context"
 	"crypto/tls"
-	"fmt"
-	"math/rand"
-	"os"
 	"runtime/debug"
 	"sync"
 	"time"
 
 	"github.com/pingcap/errors"
+	"github.com/pingcap/failpoint"
 	"github.com/pingcap/log"
 	preparesnap "github.com/pingcap/tidb/br/pkg/backup/prepare_snap"
 	berrors "github.com/pingcap/tidb/br/pkg/errors"
@@ -138,12 +136,19 @@ func AdaptEnvForSnapshotBackup(ctx context.Context, cfg *PauseGcConfig) error {
 	cx.run(func() error { return pauseGCKeeper(cx) })
 	cx.run(func() error {
 		log.Info("Pause scheduler waiting all connections established.")
-		<-initChan
+		select {
+		case <-initChan:
+		case <-cx.Done():
+			return cx.Err()
+		}
 		log.Info("Pause scheduler noticed connections established.")
 		return pauseSchedulerKeeper(cx)
 	})
 	cx.run(func() error { return pauseAdminAndWaitApply(cx, initChan) })
 	go func() {
+		failpoint.Inject("SkipReadyHint", func() {
+			failpoint.Return()
+		})
 		cx.rdGrp.Wait()
 		if cfg.OnAllReady != nil {
 			cfg.OnAllReady()
@@ -192,14 +197,6 @@ func pauseAdminAndWaitApply(cx *AdaptEnvForSnapshotBackupContext, afterConnectio
 	return nil
 }
 
-func getCallerName() string {
-	name, err := os.Hostname()
-	if err != nil {
-		name = fmt.Sprintf("UNKNOWN-%d", rand.Int63())
-	}
-	return fmt.Sprintf("operator@%sT%d#%d", name, time.Now().Unix(), os.Getpid())
-}
-
 func pauseGCKeeper(cx *AdaptEnvForSnapshotBackupContext) (err error) {
 	// Note: should we remove the service safepoint as soon as this exits?
 	sp := utils.BRServiceSafePoint{

tests/realtikvtest/brietest/operator_test.go

@@ -23,6 +23,7 @@ import (
 	"time"
 
 	"github.com/google/uuid"
+	"github.com/pingcap/failpoint"
 	"github.com/pingcap/kvproto/pkg/import_sstpb"
 	"github.com/pingcap/kvproto/pkg/kvrpcpb"
 	"github.com/pingcap/tidb/br/pkg/task"
@@ -224,3 +225,33 @@ func TestOperator(t *testing.T) {
 	verifySchedulerNotStopped(req, cfg)
 	verifyGCNotStopped(req, cfg)
 }
+
+func TestFailure(t *testing.T) {
+	req := require.New(t)
+	req.NoError(failpoint.Enable("github.com/pingcap/tidb/br/pkg/backup/prepare_snap/PrepareConnectionsErr", "return()"))
+	// Make goleak happy.
+	req.NoError(failpoint.Enable("github.com/pingcap/tidb/br/pkg/task/operator/SkipReadyHint", "return()"))
+	defer func() {
+		req.NoError(failpoint.Disable("github.com/pingcap/tidb/br/pkg/backup/prepare_snap/PrepareConnectionsErr"))
+		req.NoError(failpoint.Disable("github.com/pingcap/tidb/br/pkg/task/operator/SkipReadyHint"))
+	}()
+
+	cfg := operator.PauseGcConfig{
+		Config: task.Config{
+			PD: []string{"127.0.0.1:2379"},
+		},
+		TTL:       5 * time.Minute,
+		SafePoint: oracle.GoTimeToTS(time.Now()),
+	}
+
+	verifyGCNotStopped(req, cfg)
+	verifySchedulerNotStopped(req, cfg)
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	err := operator.AdaptEnvForSnapshotBackup(ctx, &cfg)
+	require.Error(t, err)
+
+	verifyGCNotStopped(req, cfg)
+	verifySchedulerNotStopped(req, cfg)
+}