lightning: fix panic when nextKey twice (#40959) (#41017)

close #40934

br/pkg/lightning/backend/local/local.go

@@ -1385,6 +1385,9 @@ func (local *local) ImportEngine(ctx context.Context, engineUUID uuid.UUID, regi
 		// the table when table is created.
 		needSplit := len(unfinishedRanges) > 1 || lfTotalSize > regionSplitSize || lfLength > regionSplitKeys
 		// split region by given ranges
+		failpoint.Inject("failToSplit", func(_ failpoint.Value) {
+			needSplit = true
+		})
 		for i := 0; i < maxRetryTimes; i++ {
 			err = local.SplitAndScatterRegionInBatches(ctx, unfinishedRanges, lf.tableInfo, needSplit, regionSplitSize, maxBatchSplitRanges)
 			if err == nil || common.IsContextCanceledError(err) {
@@ -1856,7 +1859,8 @@ func nextKey(key []byte) []byte {
 
 	// in tikv <= 4.x, tikv will truncate the row key, so we should fetch the next valid row key
 	// See: https://github.com/tikv/tikv/blob/f7f22f70e1585d7ca38a59ea30e774949160c3e8/components/raftstore/src/coprocessor/split_observer.rs#L36-L41
-	if tablecodec.IsRecordKey(key) {
+	// we only do this for IntHandle, which is checked by length
+	if tablecodec.IsRecordKey(key) && len(key) == tablecodec.RecordRowKeyLen {
 		tableID, handle, _ := tablecodec.DecodeRecordKey(key)
 		nextHandle := handle.Next()
 		// int handle overflow, use the next table prefix as nextKey
@@ -1866,7 +1870,7 @@ func nextKey(key []byte) []byte {
 		return tablecodec.EncodeRowKeyWithHandle(tableID, nextHandle)
 	}
 
-	// if key is an index, directly append a 0x00 to the key.
+	// for index key and CommonHandle, directly append a 0x00 to the key.
 	res := make([]byte, 0, len(key)+1)
 	res = append(res, key...)
 	res = append(res, 0)

br/pkg/lightning/backend/local/local_test.go

@@ -116,10 +116,21 @@ func TestNextKey(t *testing.T) {
 		require.NoError(t, err)
 		nextHdl, err := tidbkv.NewCommonHandle(nextKeyBytes)
 		require.NoError(t, err)
-		expectNextKey := []byte(tablecodec.EncodeRowKeyWithHandle(1, nextHdl))
-		require.Equal(t, expectNextKey, nextKey(key))
+		nextValidKey := []byte(tablecodec.EncodeRowKeyWithHandle(1, nextHdl))
+		// nextKey may return a key that can't be decoded, but it must not be larger than the valid next key.
+		require.True(t, bytes.Compare(nextKey(key), nextValidKey) <= 0, "datums: %v", datums)
 	}
 
+	// a special case that when len(string datum) % 8 == 7, nextKey twice should not panic.
+	keyBytes, err := codec.EncodeKey(stmtCtx, nil, types.NewStringDatum("1234567"))
+	require.NoError(t, err)
+	h, err := tidbkv.NewCommonHandle(keyBytes)
+	require.NoError(t, err)
+	key = tablecodec.EncodeRowKeyWithHandle(1, h)
+	nextOnce := nextKey(key)
+	// should not panic
+	_ = nextKey(nextOnce)
+
 	// dIAAAAAAAAD/PV9pgAAAAAD/AAABA4AAAAD/AAAAAQOAAAD/AAAAAAEAAAD8
 	// a index key with: table: 61, index: 1, int64: 1, int64: 1
 	a := []byte{116, 128, 0, 0, 0, 0, 0, 0, 255, 61, 95, 105, 128, 0, 0, 0, 0, 255, 0, 0, 1, 3, 128, 0, 0, 0, 255, 0, 0, 0, 1, 3, 128, 0, 0, 255, 0, 0, 0, 0, 1, 0, 0, 0, 252}

br/pkg/lightning/backend/local/localhelper.go

@@ -27,6 +27,7 @@ import (
 
 	"github.com/docker/go-units"
 	"github.com/pingcap/errors"
+	"github.com/pingcap/failpoint"
 	sst "github.com/pingcap/kvproto/pkg/import_sstpb"
 	"github.com/pingcap/kvproto/pkg/metapb"
 	"github.com/pingcap/kvproto/pkg/pdpb"
@@ -381,7 +382,14 @@ func fetchTableRegionSizeStats(ctx context.Context, db *sql.DB, tableID int64) (
 	return stats, errors.Trace(err)
 }
 
-func (local *local) BatchSplitRegions(ctx context.Context, region *split.RegionInfo, keys [][]byte) (*split.RegionInfo, []*split.RegionInfo, error) {
+func (local *local) BatchSplitRegions(
+	ctx context.Context,
+	region *split.RegionInfo,
+	keys [][]byte,
+) (*split.RegionInfo, []*split.RegionInfo, error) {
+	failpoint.Inject("failToSplit", func(_ failpoint.Value) {
+		failpoint.Return(nil, nil, errors.New("retryable error"))
+	})
 	region, newRegions, err := local.splitCli.BatchSplitRegionsWithOrigin(ctx, region, keys)
 	if err != nil {
 		return nil, nil, errors.Annotatef(err, "batch split regions failed")

br/tests/lightning_local_backend/data/cpeng.a-schema.sql

@@ -1 +1 @@
-create table a (c int);
+create table a (c VARCHAR(20) PRIMARY KEY);

br/tests/lightning_local_backend/data/cpeng.a.1.sql

@@ -1 +1 @@
-insert into a values (1);
+insert into a values ('0000001');

br/tests/lightning_local_backend/data/cpeng.a.2.sql

@@ -1 +1 @@
-insert into a values (2);
+insert into a values ('0000002');

br/tests/lightning_local_backend/data/cpeng.a.3.sql

@@ -1 +1 @@
-insert into a values (3),(4);
+insert into a values ('0000003'),('0000004');

br/tests/lightning_local_backend/run.sh

@@ -34,10 +34,11 @@ grep -Fq 'table(s) [`cpeng`.`a`, `cpeng`.`b`] are not empty' $TEST_DIR/lightning
 
 
 # First, verify that inject with not leader error is fine.
-export GO_FAILPOINTS='github.com/pingcap/tidb/br/pkg/lightning/backend/local/FailIngestMeta=1*return("notleader")'
+export GO_FAILPOINTS='github.com/pingcap/tidb/br/pkg/lightning/backend/local/FailIngestMeta=1*return("notleader");github.com/pingcap/tidb/br/pkg/lightning/backend/local/failToSplit=2*return("")'
 rm -f "$TEST_DIR/lightning-local.log"
 run_sql 'DROP DATABASE IF EXISTS cpeng;'
-run_lightning --backend local --enable-checkpoint=1 --log-file "$TEST_DIR/lightning-local.log" --config "tests/$TEST_NAME/config.toml"
+run_lightning --backend local --enable-checkpoint=1 --log-file "$TEST_DIR/lightning-local.log" --config "tests/$TEST_NAME/config.toml" -L debug
+grep -Eq "split regions.*retryable error" "$TEST_DIR/lightning-local.log"
 
 # Check that everything is correctly imported
 run_sql 'SELECT count(*), sum(c) FROM cpeng.a'

kv/key.go

@@ -140,6 +140,7 @@ type Handle interface {
 	// IntValue returns the int64 value if IsInt is true, it panics if IsInt returns false.
 	IntValue() int64
 	// Next returns the minimum handle that is greater than this handle.
+	// The returned handle is not guaranteed to be able to decode.
 	Next() Handle
 	// Equal returns if the handle equals to another handle, it panics if the types are different.
 	Equal(h Handle) bool
@@ -279,6 +280,7 @@ func (ch *CommonHandle) IntValue() int64 {
 }
 
 // Next implements the Handle interface.
+// Note that the returned encoded field is not guaranteed to be able to decode.
 func (ch *CommonHandle) Next() Handle {
 	return &CommonHandle{
 		encoded:       Key(ch.encoded).PrefixNext(),