Ensure that the metadata in the TOC matches the tar-split

Signed-off-by: Miloslav Trmač <[email protected]>

go.mod

@@ -30,6 +30,7 @@ require (
 	github.com/tchap/go-patricia/v2 v2.3.1
 	github.com/ulikunitz/xz v0.5.12
 	github.com/vbatts/tar-split v0.11.5
+	golang.org/x/exp v0.0.0-20231006140011-7918f672742d
 	golang.org/x/sys v0.22.0
 	gotest.tools v2.2.0+incompatible
 )

go.sum

@@ -119,6 +119,8 @@ go.opencensus.io v0.24.0/go.mod h1:vNK8G9p7aAivkbmorf4v+7Hgx+Zs0yY+0fOtgBfjQKo=
 golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACkg1iLfiJU5Ep61QUkGW8qpdssI0+w=
 golang.org/x/crypto v0.0.0-20200622213623-75b288015ac9/go.mod h1:LzIPMQfyMNhhGPhUkYOs5KpL4U8rLKemX1yGLhDgUto=
 golang.org/x/exp v0.0.0-20190121172915-509febef88a4/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
+golang.org/x/exp v0.0.0-20231006140011-7918f672742d h1:jtJma62tbqLibJ5sFQz8bKtEM8rJBtfilJ2qTU199MI=
+golang.org/x/exp v0.0.0-20231006140011-7918f672742d/go.mod h1:ldy0pHrwJyGW56pPQzzkH36rKxoZW1tw7ZJpeKx+hdo=
 golang.org/x/lint v0.0.0-20181026193005-c67002cb31c3/go.mod h1:UVdnD1Gm6xHRNCYTkRU2/jEulfH38KcIWyp/GAMgvoE=
 golang.org/x/lint v0.0.0-20190227174305-5b3e6a55c961/go.mod h1:wehouNa3lNwaWXcvxsM5YxQ5yQlVC4a0KAMCusXpPoU=
 golang.org/x/lint v0.0.0-20190313153728-d0100b6bd8b3/go.mod h1:6SW0HCj/g11FgYtHlgUYUwCkIfeOF89ocIRzGO/8vkc=

pkg/chunked/compression_linux.go

@@ -5,13 +5,16 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"maps"
 	"strconv"
+	"time"
 
 	"github.com/containers/storage/pkg/chunked/internal"
 	"github.com/klauspost/compress/zstd"
 	"github.com/klauspost/pgzip"
 	digest "github.com/opencontainers/go-digest"
 	"github.com/vbatts/tar-split/archive/tar"
+	expMaps "golang.org/x/exp/maps"
 )
 
 var typesToTar = map[string]byte{
@@ -221,6 +224,12 @@ func readZstdChunkedManifest(blobStream ImageSourceSeekable, tocDigest digest.Di
 		if err != nil {
 			return nil, nil, nil, 0, fmt.Errorf("validating and decompressing tar-split: %w", err)
 		}
+		// We use the TOC for creating on-disk files, but the tar-split for creating metadata
+		// when exporting the layer contents. Ensure the two match, otherwise local inspection of a container
+		// might be misleading about the exported contents.
+		if err := ensureTOCMatchesTarSplit(toc, decodedTarSplit); err != nil {
+			return nil, nil, nil, 0, fmt.Errorf("tar-split and TOC data is inconsistent: %w", err)
+		}
 	} else if tarSplitChunk.Offset > 0 {
 		// We must ignore the tar-split when the digest is not present in the TOC, because we can’t authenticate it.
 		//
@@ -234,6 +243,121 @@ func readZstdChunkedManifest(blobStream ImageSourceSeekable, tocDigest digest.Di
 	return decodedBlob, toc, decodedTarSplit, int64(manifestChunk.Offset), err
 }
 
+// ensureTOCMatchesTarSplit validates that toc and tarSplit contain _exactly_ the same entries.
+func ensureTOCMatchesTarSplit(toc *internal.TOC, tarSplit []byte) error {
+	pendingFiles := map[string]*internal.FileMetadata{} // Name -> an entry in toc.Entries
+	for i := range toc.Entries {
+		e := &toc.Entries[i]
+		if e.Type != internal.TypeChunk {
+			if _, ok := pendingFiles[e.Name]; ok {
+				return fmt.Errorf("TOC contains duplicate entries for path %q", e.Name)
+			}
+			pendingFiles[e.Name] = e
+		}
+	}
+
+	if err := iterateTarSplit(tarSplit, func(hdr *tar.Header) error {
+		e, ok := pendingFiles[hdr.Name]
+		if !ok {
+			return fmt.Errorf("tar-split contains an entry for %q missing in TOC", hdr.Name)
+		}
+		delete(pendingFiles, hdr.Name)
+		expected, err := internal.NewFileMetadata(hdr)
+		if err != nil {
+			return fmt.Errorf("determining expected metadata for %q: %w", hdr.Name, err)
+		}
+		if err := ensureFileMetadataAttributesMatch(e, &expected); err != nil {
+			return fmt.Errorf("TOC and tar-split metadata doesn’t match: %w", err)
+		}
+
+		return nil
+	}); err != nil {
+		return err
+	}
+	if len(pendingFiles) != 0 {
+		remaining := expMaps.Keys(pendingFiles)
+		if len(remaining) > 5 {
+			remaining = remaining[:5] // Just to limit the size of the output.
+		}
+		return fmt.Errorf("TOC contains entries not present in tar-split, incl. %q", remaining)
+	}
+	return nil
+}
+
+// ensureTimePointersMatch ensures that a and b are equal
+func ensureTimePointersMatch(a, b *time.Time) error {
+	switch {
+	case a == nil && b == nil:
+		return nil
+	case a == nil:
+		return fmt.Errorf("nil != %v", *b)
+	case b == nil:
+		return fmt.Errorf("%v != nil", *a)
+	default:
+		if a.Equal(*b) {
+			return nil
+		}
+		return fmt.Errorf("%v != %v", *a, *b)
+	}
+}
+
+// ensureFileMetadataAttributesMatch ensures that a and b match in file attributes (it ignores entries relevant to locating data
+// in the tar stream or matching contents)
+func ensureFileMetadataAttributesMatch(a, b *internal.FileMetadata) error {
+	// Keep this in sync with internal.FileMetadata!
+
+	if a.Type != b.Type {
+		return fmt.Errorf("mismatch of Type: %q != %q", a.Type, b.Type)
+	}
+	if a.Name != b.Name {
+		return fmt.Errorf("mismatch of Name: %q != %q", a.Name, b.Name)
+	}
+	if a.Linkname != b.Linkname {
+		return fmt.Errorf("mismatch of Linkname: %q != %q", a.Linkname, b.Linkname)
+	}
+	if a.Mode != b.Mode {
+		return fmt.Errorf("mismatch of Mode: %q != %q", a.Mode, b.Mode)
+	}
+	if a.Size != b.Size {
+		return fmt.Errorf("mismatch of Size: %q != %q", a.Size, b.Size)
+	}
+	if a.UID != b.UID {
+		return fmt.Errorf("mismatch of UID: %q != %q", a.UID, b.UID)
+	}
+	if a.GID != b.GID {
+		return fmt.Errorf("mismatch of GID: %q != %q", a.GID, b.GID)
+	}
+
+	if err := ensureTimePointersMatch(a.ModTime, b.ModTime); err != nil {
+		return fmt.Errorf("mismatch of ModTime: %w", err)
+	}
+	if err := ensureTimePointersMatch(a.AccessTime, b.AccessTime); err != nil {
+		return fmt.Errorf("mismatch of AccessTime: %w", err)
+	}
+	if err := ensureTimePointersMatch(a.ChangeTime, b.ChangeTime); err != nil {
+		return fmt.Errorf("mismatch of ChangeTime: %w", err)
+	}
+	if a.Devmajor != b.Devmajor {
+		return fmt.Errorf("mismatch of Devmajor: %q != %q", a.Devmajor, b.Devmajor)
+	}
+	if a.Devminor != b.Devminor {
+		return fmt.Errorf("mismatch of Devminor: %q != %q", a.Devminor, b.Devminor)
+	}
+	if !maps.Equal(a.Xattrs, b.Xattrs) {
+		return fmt.Errorf("mismatch of Xattrs: %q != %q", a.Xattrs, b.Xattrs)
+	}
+
+	// Digest is not compared
+	// Offset is not compared
+	// EndOffset is not compared
+
+	// ChunkSize is not compared
+	// ChunkOffset is not compared
+	// ChunkDigest is not compared
+	// ChunkType is not compared
+	return nil
+}
+
 func decodeAndValidateBlob(blob []byte, lengthUncompressed uint64, expectedCompressedChecksum string) ([]byte, error) {
 	d, err := digest.Parse(expectedCompressedChecksum)
 	if err != nil {

pkg/chunked/internal/compression.go

@@ -43,6 +43,8 @@ type TOC struct {
 // is used instead of that in the tar stream.  The contents of the tar stream
 // are not used in this scenario.
 type FileMetadata struct {
+	// If you add any fields, update ensureFileMetadataMatches as well!
+
 	// The metadata below largely duplicates that in the tar headers.
 	Type       string            `json:"type"`
 	Name       string            `json:"name"`

pkg/chunked/tar_split_linux.go

@@ -0,0 +1,68 @@
+package chunked
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+
+	"github.com/vbatts/tar-split/archive/tar"
+	"github.com/vbatts/tar-split/tar/storage"
+)
+
+// iterateTarSplit calls handler for each tar header in tarSplit
+func iterateTarSplit(tarSplit []byte, handler func(hdr *tar.Header) error) error {
+	// This, strictly speaking, hard-codes undocumented assumptions about how github.com/vbatts/tar-split/tar/asm.NewInputTarStream
+	// forms the tar-split contents. Pragmatically, NewInputTarStream should always produce storage.FileType entries at least
+	// for every non-empty file, which constraints it basically to the output we expect.
+	//
+	// Specifically, we assume:
+	// - There is a separate SegmentType entry for every tar header, but only one SegmentType entry for the full header incl. any extensions
+	// - (There is a FileType entry for every tar header, we ignore it)
+	// - Trailing padding of a file, if any, is included in the next SegmentType entry
+	// - At the end, there may be SegmentType entries just for the terminating zero blocks.
+
+	unpacker := storage.NewJSONUnpacker(bytes.NewReader(tarSplit))
+	for {
+		tsEntry, err := unpacker.Next()
+		if err != nil {
+			if err == io.EOF {
+				return nil
+			}
+			return fmt.Errorf("reading tar-split entries: %w", err)
+		}
+		switch tsEntry.Type {
+		case storage.SegmentType:
+			payload := tsEntry.Payload
+			// This is horrible, but we don’t know how much padding to skip. (It can be computed from the previous hdr.Size for non-sparse
+			// files, but for sparse files that is set to the logical size.)
+			//
+			// First, assume that all padding is zero bytes.
+			// A tar header starts with a file name, which might in principle be empty, but
+			// at least https://github.com/opencontainers/image-spec/blob/main/layer.md#populate-initial-filesystem suggests that
+			// the tar name should never be empty (it should be ".", or maybe "./").
+			//
+			// This will cause us to skip all zero bytes in the trailing blocks, but that’s fine.
+			i := 0
+			for i < len(payload) && payload[i] == 0 {
+				i++
+			}
+			payload = payload[i:]
+			tr := tar.NewReader(bytes.NewReader(payload))
+			hdr, err := tr.Next()
+			if err != nil {
+				if err == io.EOF { // Probably the last entry, but let’s let the unpacker drive that.
+					break
+				}
+				return fmt.Errorf("decoding a tar header from a tar-split entry: %w", err)
+			}
+			if err := handler(hdr); err != nil {
+				return err
+			}
+
+		case storage.FileType:
+			// Nothing
+		default:
+			return fmt.Errorf("unexpected tar-split entry type %q", tsEntry.Type)
+		}
+	}
+}

pkg/chunked/tar_split_linux_test.go

@@ -0,0 +1,104 @@
+package chunked
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+	"testing"
+	"time"
+
+	"github.com/containers/storage/pkg/chunked/internal"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"github.com/vbatts/tar-split/archive/tar"
+	"github.com/vbatts/tar-split/tar/asm"
+	"github.com/vbatts/tar-split/tar/storage"
+)
+
+func createTestTarheader(index int, typeFlag byte, size int64) tar.Header {
+	n := (index + 1) * 100 // Use predictable, but distinct, values for all headers
+
+	res := tar.Header{
+		Typeflag:   typeFlag,
+		Name:       fmt.Sprintf("name%d", n),
+		Size:       size,
+		Mode:       int64(n + 1),
+		Uid:        n + 2,
+		Gid:        n + 3,
+		Uname:      fmt.Sprintf("user%d", n),
+		Gname:      fmt.Sprintf("group%d", n),
+		ModTime:    time.Unix(int64(n+4), 0),
+		AccessTime: time.Unix(int64(n+5), 0),
+		ChangeTime: time.Unix(int64(n+6), 0),
+		PAXRecords: map[string]string{fmt.Sprintf("key%d", n): fmt.Sprintf("value%d", n)},
+		Format:     tar.FormatPAX, // We must set a format, in the default one AccessTime and ChangeTime are discarded.
+	}
+	switch res.Typeflag {
+	case tar.TypeLink, tar.TypeSymlink:
+		res.Linkname = fmt.Sprintf("link%d", n)
+	case tar.TypeChar, tar.TypeBlock:
+		res.Devmajor = int64(n + 7)
+		res.Devminor = int64(n + 8)
+	}
+	return res
+}
+
+func TestIterateTarSplit(t *testing.T) {
+	entries := []struct {
+		typeFlag byte
+		size     int64
+	}{
+		{tar.TypeReg, 0},
+		{tar.TypeReg, 1},
+		{tar.TypeReg, 511},
+		{tar.TypeReg, 512},
+		{tar.TypeReg, 513},
+		{tar.TypeLink, 0},
+		{tar.TypeSymlink, 0},
+		{tar.TypeChar, 0},
+		{tar.TypeBlock, 0},
+		{tar.TypeDir, 0},
+		{tar.TypeFifo, 0},
+	}
+
+	var tarball bytes.Buffer
+	var expected []tar.Header
+	w := tar.NewWriter(&tarball)
+	for i, e := range entries {
+		hdr := createTestTarheader(i, e.typeFlag, e.size)
+		err := w.WriteHeader(&hdr)
+		require.NoError(t, err)
+		data := make([]byte, e.size)
+		_, err = w.Write(data)
+		require.NoError(t, err)
+		expected = append(expected, hdr)
+	}
+	err := w.Close()
+	require.NoError(t, err)
+
+	var tarSplit bytes.Buffer
+	tsReader, err := asm.NewInputTarStream(&tarball, storage.NewJSONPacker(&tarSplit), storage.NewDiscardFilePutter())
+	require.NoError(t, err)
+	_, err = io.Copy(io.Discard, tsReader)
+	require.NoError(t, err)
+
+	var actual []tar.Header
+	err = iterateTarSplit(tarSplit.Bytes(), func(hdr *tar.Header) error {
+		actual = append(actual, *hdr)
+		return nil
+	})
+	require.NoError(t, err)
+
+	assert.Equal(t, len(expected), len(actual))
+	for i := range expected {
+		// We would have to open-code an equality comparison of time.Time values; instead, convert to FileMetadata,
+		// because we already have that implemented for that type — and because it provides a tiny bit of code coverage
+		// testing for ensureFileMetadataAttributesMatch.
+		expected1, err := internal.NewFileMetadata(&expected[i])
+		require.NoError(t, err, i)
+		actual1, err := internal.NewFileMetadata(&actual[i])
+		require.NoError(t, err, i)
+		err = ensureFileMetadataAttributesMatch(&expected1, &actual1)
+		assert.NoError(t, err, i)
+	}
+}