diff --git a/br/pkg/lightning/mydump/csv_parser.go b/br/pkg/lightning/mydump/csv_parser.go
index 61cc71616ceb1..2094d66652aea 100644
--- a/br/pkg/lightning/mydump/csv_parser.go
+++ b/br/pkg/lightning/mydump/csv_parser.go
@@ -344,7 +344,7 @@ func (parser *CSVParser) readUntil(chars *byteSet) ([]byte, byte, error) {
 	var buf []byte
 	for {
 		buf = append(buf, parser.buf...)
-		if len(buf) > LargestEntryLimit {
+		if parser.checkRowLen && parser.pos-parser.rowStartPos+int64(len(buf)) > int64(LargestEntryLimit) {
 			return buf, 0, errors.New("size of row cannot exceed the max value of txn-entry-size-limit")
 		}
 		parser.buf = nil
@@ -513,6 +513,8 @@ func (parser *CSVParser) replaceEOF(err error, replaced error) error {
 
 // ReadRow reads a row from the datafile.
 func (parser *CSVParser) ReadRow() error {
+	parser.beginRowLenCheck()
+	defer parser.endRowLenCheck()
 	row := &parser.lastRow
 	row.Length = 0
 	row.RowID++
@@ -563,6 +565,8 @@ func (parser *CSVParser) ReadRow() error {
 
 // ReadColumns reads the columns of this CSV file.
 func (parser *CSVParser) ReadColumns() error {
+	parser.beginRowLenCheck()
+	defer parser.endRowLenCheck()
 	columns, err := parser.readRecord(nil)
 	if err != nil {
 		return errors.Trace(err)
@@ -582,6 +586,8 @@ func (parser *CSVParser) ReadColumns() error {
 // returns the file offset beyond the terminator.
 // This function is used in strict-format dividing a CSV file.
 func (parser *CSVParser) ReadUntilTerminator() (int64, error) {
+	parser.beginRowLenCheck()
+	defer parser.endRowLenCheck()
 	for {
 		_, firstByte, err := parser.readUntil(&parser.newLineByteSet)
 		if err != nil {
diff --git a/br/pkg/lightning/mydump/csv_parser_test.go b/br/pkg/lightning/mydump/csv_parser_test.go
index da06c15ed39d9..da76c7157d03b 100644
--- a/br/pkg/lightning/mydump/csv_parser_test.go
+++ b/br/pkg/lightning/mydump/csv_parser_test.go
@@ -688,20 +688,49 @@ func TestTooLargeRow(t *testing.T) {
 			Delimiter: `"`,
 		},
 	}
-	var testCase bytes.Buffer
-	testCase.WriteString("a,b,c,d")
-	// WARN: will take up 10KB memory here.
-	mydump.LargestEntryLimit = 10 * 1024
-	for i := 0; i < mydump.LargestEntryLimit; i++ {
-		testCase.WriteByte('d')
-	}
-	charsetConvertor, err := mydump.NewCharsetConvertor(cfg.DataCharacterSet, cfg.DataInvalidCharReplace)
-	require.NoError(t, err)
-	parser, err := mydump.NewCSVParser(context.Background(), &cfg.CSV, mydump.NewStringReader(testCase.String()), int64(config.ReadBlockSize), ioWorkers, false, charsetConvertor)
-	require.NoError(t, err)
-	e := parser.ReadRow()
-	require.Error(t, e)
-	require.Contains(t, e.Error(), "size of row cannot exceed the max value of txn-entry-size-limit")
+	bak := mydump.LargestEntryLimit
+	t.Cleanup(func() {
+		mydump.LargestEntryLimit = bak
+	})
+	mydump.LargestEntryLimit = 1024
+	t.Run("too long field", func(t *testing.T) {
+		var dataBuf bytes.Buffer
+		dataBuf.WriteString("a,b,c,d")
+		for i := 0; i < mydump.LargestEntryLimit; i++ {
+			dataBuf.WriteByte('d')
+		}
+		require.Greater(t, dataBuf.Len(), mydump.LargestEntryLimit)
+		charsetConvertor, err := mydump.NewCharsetConvertor(cfg.DataCharacterSet, cfg.DataInvalidCharReplace)
+		require.NoError(t, err)
+		parser, err := mydump.NewCSVParser(context.Background(), &cfg.CSV, mydump.NewStringReader(dataBuf.String()), int64(config.ReadBlockSize), ioWorkers, false, charsetConvertor)
+		require.NoError(t, err)
+		e := parser.ReadRow()
+		require.Error(t, e)
+		require.Contains(t, e.Error(), "size of row cannot exceed the max value of txn-entry-size-limit")
+	})
+
+	t.Run("field is short, but whole row too long", func(t *testing.T) {
+		var dataBuf bytes.Buffer
+		for i := 0; i < 16; i++ {
+			if i > 0 {
+				dataBuf.WriteByte(',')
+			}
+			for j := 0; j < mydump.LargestEntryLimit/16; j++ {
+				dataBuf.WriteByte('d')
+			}
+		}
+		for i := 0; i < mydump.LargestEntryLimit-dataBuf.Len()+16; i++ {
+			dataBuf.WriteByte('d')
+		}
+		require.Greater(t, dataBuf.Len(), mydump.LargestEntryLimit)
+		charsetConvertor, err := mydump.NewCharsetConvertor(cfg.DataCharacterSet, cfg.DataInvalidCharReplace)
+		require.NoError(t, err)
+		parser, err := mydump.NewCSVParser(context.Background(), &cfg.CSV, mydump.NewStringReader(dataBuf.String()), int64(config.ReadBlockSize), ioWorkers, false, charsetConvertor)
+		require.NoError(t, err)
+		e := parser.ReadRow()
+		require.Error(t, e)
+		require.Contains(t, e.Error(), "size of row cannot exceed the max value of txn-entry-size-limit")
+	})
 }
 
 func TestSpecialChars(t *testing.T) {
diff --git a/br/pkg/lightning/mydump/parser.go b/br/pkg/lightning/mydump/parser.go
index 0ac82ce189d71..53d478d2749f4 100644
--- a/br/pkg/lightning/mydump/parser.go
+++ b/br/pkg/lightning/mydump/parser.go
@@ -54,11 +54,17 @@ type blockParser struct {
 
 	// cache
 	remainBuf *bytes.Buffer
+	// holds the cached parsable data after last readBlock. all data inside is
+	// unparsed at the moment of the return of last readBlock, for current unparsed
+	// data, use buf.
 	appendBuf *bytes.Buffer
 
 	// the Logger associated with this parser for reporting failure
 	Logger  log.Logger
 	metrics *metric.Metrics
+
+	checkRowLen bool
+	rowStartPos int64
 }
 
 func makeBlockParser(
@@ -164,6 +170,15 @@ func NewChunkParser(
 	}
 }
 
+func (parser *blockParser) beginRowLenCheck() {
+	parser.checkRowLen = true
+	parser.rowStartPos = parser.pos
+}
+
+func (parser *blockParser) endRowLenCheck() {
+	parser.checkRowLen = false
+}
+
 // SetPos changes the reported position and row ID.
 func (parser *blockParser) SetPos(pos int64, rowID int64) error {
 	p, err := parser.reader.Seek(pos, io.SeekStart)