advancer: fix the incorrect gc safepoint behaviours (#52835) (#56698)

ref #52082

Author: ti-chi-bot

br/pkg/streamhelper/advancer.go

@@ -453,7 +453,7 @@ func (c *CheckpointAdvancer) onTaskEvent(ctx context.Context, e TaskEvent) error
 		if err := c.env.ClearV3GlobalCheckpointForTask(ctx, e.Name); err != nil {
 			log.Warn("failed to clear global checkpoint", logutil.ShortError(err))
 		}
-		if _, err := c.env.BlockGCUntil(ctx, 0); err != nil {
+		if err := c.env.UnblockGC(ctx); err != nil {
 			log.Warn("failed to remove service GC safepoint", logutil.ShortError(err))
 		}
 		metrics.LastCheckpoint.DeleteLabelValues(e.Name)

br/pkg/streamhelper/advancer_env.go

@@ -4,8 +4,10 @@ package streamhelper
 
 import (
 	"context"
+	"math"
 	"time"
 
+	"github.com/pingcap/errors"
 	logbackup "github.com/pingcap/kvproto/pkg/logbackuppb"
 	"github.com/pingcap/tidb/br/pkg/utils"
 	"github.com/pingcap/tidb/pkg/config"
@@ -46,7 +48,21 @@ type PDRegionScanner struct {
 // Returns the minimal service GC safe point across all services.
 // If the arguments is `0`, this would remove the service safe point.
 func (c PDRegionScanner) BlockGCUntil(ctx context.Context, at uint64) (uint64, error) {
-	return c.UpdateServiceGCSafePoint(ctx, logBackupServiceID, int64(logBackupSafePointTTL.Seconds()), at)
+	minimalSafePoint, err := c.UpdateServiceGCSafePoint(
+		ctx, logBackupServiceID, int64(logBackupSafePointTTL.Seconds()), at)
+	if err != nil {
+		return 0, errors.Annotate(err, "failed to block gc until")
+	}
+	if minimalSafePoint > at {
+		return 0, errors.Errorf("minimal safe point %d is greater than the target %d", minimalSafePoint, at)
+	}
+	return at, nil
+}
+
+func (c PDRegionScanner) UnblockGC(ctx context.Context) error {
+	// set ttl to 0, means remove the safe point.
+	_, err := c.UpdateServiceGCSafePoint(ctx, logBackupServiceID, 0, math.MaxUint64)
+	return err
 }
 
 // TODO: It should be able to synchoronize the current TS with the PD.

br/pkg/streamhelper/advancer_test.go

@@ -214,7 +214,7 @@ func TestGCServiceSafePoint(t *testing.T) {
 	req.Eventually(func() bool {
 		env.fakeCluster.mu.Lock()
 		defer env.fakeCluster.mu.Unlock()
-		return env.serviceGCSafePoint == 0
+		return env.serviceGCSafePoint != 0 && env.serviceGCSafePointDeleted
 	}, 3*time.Second, 100*time.Millisecond)
 }
 

br/pkg/streamhelper/basic_lib_for_test.go

@@ -102,10 +102,11 @@ type fakeCluster struct {
 	maxTs     uint64
 	testCtx   *testing.T
 
-	onGetClient        func(uint64) error
-	onClearCache       func(uint64) error
-	serviceGCSafePoint uint64
-	currentTS          uint64
+	onGetClient               func(uint64) error
+	onClearCache              func(uint64) error
+	serviceGCSafePoint        uint64
+	serviceGCSafePointDeleted bool
+	currentTS                 uint64
 }
 
 func (r *region) splitAt(newID uint64, k string) *region {
@@ -266,17 +267,20 @@ func (f *fakeStore) GetLastFlushTSOfRegion(ctx context.Context, in *logbackup.Ge
 func (f *fakeCluster) BlockGCUntil(ctx context.Context, at uint64) (uint64, error) {
 	f.mu.Lock()
 	defer f.mu.Unlock()
-	if at == 0 {
-		f.serviceGCSafePoint = at
-		return at, nil
-	}
 	if f.serviceGCSafePoint > at {
 		return f.serviceGCSafePoint, nil
 	}
 	f.serviceGCSafePoint = at
 	return at, nil
 }
 
+func (f *fakeCluster) UnblockGC(ctx context.Context) error {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	f.serviceGCSafePointDeleted = true
+	return nil
+}
+
 func (f *fakeCluster) FetchCurrentTS(ctx context.Context) (uint64, error) {
 	return f.currentTS, nil
 }

br/pkg/streamhelper/regioniter.go

@@ -43,6 +43,9 @@ type TiKVClusterMeta interface {
 	// For now, all tasks (exactly one task in fact) use the same checkpoint.
 	BlockGCUntil(ctx context.Context, at uint64) (uint64, error)
 
+	// UnblockGC used to remove the service GC safe point in PD.
+	UnblockGC(ctx context.Context) error
+
 	FetchCurrentTS(ctx context.Context) (uint64, error)
 }
 

br/pkg/streamhelper/regioniter_test.go

@@ -83,6 +83,10 @@ func (c constantRegions) BlockGCUntil(ctx context.Context, at uint64) (uint64, e
 	return 0, status.Error(codes.Unimplemented, "Unsupported operation")
 }
 
+func (c constantRegions) UnblockGC(ctx context.Context) error {
+	return status.Error(codes.Unimplemented, "Unsupported operation")
+}
+
 // TODO: It should be able to synchoronize the current TS with the PD.
 func (c constantRegions) FetchCurrentTS(ctx context.Context) (uint64, error) {
 	return oracle.ComposeTS(time.Now().UnixMilli(), 0), nil