Read xz with a real dictionary maximum (review 65, hq issue 345)

ulikunitz/xz's ReaderConfig.DictCap is a minimum: a stream asking for a
1 GiB or 4 GiB dictionary was given it. github.com/therootcompany/xz, a port
of xz-embedded, takes a maximum and refuses a block whose LZMA2 dictionary is
larger before allocating it. 64 MiB, xz's largest preset.

internal/builder/upstream_file.go

@@ -20,7 +20,7 @@ import (
 	"time"
 
 	"github.com/novox/mesh-controller/internal/catalogue"
-	"github.com/ulikunitz/xz"
+	"github.com/therootcompany/xz"
 )
 
 // An archive whose source is somebody else's published file, pinned by URL and sha256 (novox/hq ADR 0316
@@ -194,14 +194,15 @@ func unpackUpstream(ctx context.Context, file, into, within string) error {
 	case bytes.HasPrefix(magic, []byte("BZh")):
 		stream = bzip2.NewReader(buffered)
 	case bytes.HasPrefix(magic, []byte{0xfd, '7', 'z', 'X', 'Z', 0x00}):
-		// Read in Go (github.com/ulikunitz/xz), never by a tool the build seat's image may not carry: the seat's
-		// image is the catalogue's build-agent recipe, which installs git and docker-cli alone, and a release
-		// pinned as .tar.xz failed there on 2026-10-12 (hq issue 345). The dictionary is bounded, and what it
-		// unpacks to is bounded by untar (maxUpstream, maxEntries).
-		config := xz.ReaderConfig{DictCap: xzDictCap}
-		zipped, err := config.NewReader(buffered)
+		// Read in Go (github.com/therootcompany/xz, a port of xz-embedded), never by a tool the build seat's image
+		// may not carry: the seat's image is the catalogue's build-agent recipe, which installs git and docker-cli
+		// alone, and a release pinned as .tar.xz failed there on 2026-10-12 (hq issue 345). xzDictMax is a real
+		// maximum: a block whose LZMA2 dictionary is larger is refused before its dictionary is allocated
+		// (ErrMemlimit). What the stream unpacks to is bounded by untar (maxUpstream, maxEntries).
+		zipped, err := xz.NewReader(buffered, xzDictMax)
 		if err != nil {
-			return fmt.Errorf("it is not an xz stream this build seat reads: %v", err)
+			return fmt.Errorf("it is not an xz stream this build seat reads (a dictionary of %d MiB at most): %v",
+				xzDictMax>>20, err)
 		}
 		stream = zipped
 	default:
@@ -211,9 +212,9 @@ func unpackUpstream(ctx context.Context, file, into, within string) error {
 	return untar(stream, into, within)
 }
 
-// xzDictCap is the largest xz dictionary read: xz's own presets use at most 64 MiB (-9); a stream asking for
+// xzDictMax is the largest xz dictionary read: xz's own presets use at most 64 MiB (-9); a stream asking for
 // more is refused rather than given the memory.
-const xzDictCap = 64 << 20
+const xzDictMax = 64 << 20
 
 // untar writes a tar stream's files and directories within `within` into a directory.
 func untar(stream io.Reader, into, within string) error {

internal/builder/upstream_file_test.go

@@ -6,8 +6,10 @@ import (
 	"compress/gzip"
 	"context"
 	"crypto/sha256"
+	"encoding/binary"
 	"encoding/hex"
 	"fmt"
+	"hash/crc32"
 	"io"
 	"net"
 	"net/http"
@@ -15,6 +17,7 @@ import (
 	"os"
 	"os/exec"
 	"path/filepath"
+	"runtime"
 	"sort"
 	"strings"
 	"testing"
@@ -405,3 +408,62 @@ func TestAnXzReleaseIsReadWithoutAnyXzToolOnTheBuildSeat(t *testing.T) {
 		t.Fatalf("%v", e)
 	}
 }
+
+// withDictionary is an xz stream whose first block's LZMA2 dictionary property is props, its header's CRC32
+// recomputed: a stream asking for a dictionary of its own choosing.
+func withDictionary(t *testing.T, stream []byte, props byte) []byte {
+	t.Helper()
+	out := append([]byte{}, stream...)
+	const at = 12 // the stream header
+	size := (int(out[at]) + 1) * 4
+	header := out[at : at+size]
+	flags := header[1]
+	i := 2
+	readVarint := func() {
+		for header[i]&0x80 != 0 {
+			i++
+		}
+		i++
+	}
+	if flags&0x40 != 0 {
+		readVarint()
+	}
+	if flags&0x80 != 0 {
+		readVarint()
+	}
+	if header[i] != 0x21 || header[i+1] != 0x01 {
+		t.Fatalf("the fixture's first filter is not LZMA2: % x", header)
+	}
+	header[i+2] = props
+	binary.LittleEndian.PutUint32(header[size-4:], crc32.ChecksumIEEE(header[:size-4]))
+	return out
+}
+
+func TestAnXzStreamAskingForTooLargeADictionaryIsRefused(t *testing.T) {
+	body, err := os.ReadFile(filepath.Join("testdata", "release.tar.xz"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	// 1 GiB (props 36) and 4 GiB less one (props 40): over the 64 MiB maximum, refused before allocation.
+	for _, props := range []byte{36, 40} {
+		crafted := withDictionary(t, body, props)
+		v := &vendor{files: map[string][]byte{"/r.tar.xz": crafted}}
+		base := serveVendor(t, v)
+		var before, after runtime.MemStats
+		runtime.GC()
+		runtime.ReadMemStats(&before)
+		_, err := upstreamArchive(context.Background(), t.TempDir(), catalogue.Artifact{Name: "archive", Kind: "archive",
+			URL: base + "/r.tar.xz", SHA256: sumHex(crafted), Version: "1", Within: "betterbird"}, func(string, string, ...any) {})
+		runtime.ReadMemStats(&after)
+		if err == nil || !strings.Contains(err.Error(), "dictionary") {
+			t.Errorf("props %d: a dictionary over the maximum was not refused: %v", props, err)
+		}
+		if grew := after.TotalAlloc - before.TotalAlloc; grew > 256<<20 {
+			t.Errorf("props %d: %d bytes were allocated for a refused stream", props, grew)
+		}
+	}
+	// The fixture as it is (a small dictionary) still reads.
+	if crafted := withDictionary(t, body, 0); len(crafted) == 0 {
+		t.Fatal("no fixture")
+	}
+}