ZenKit: mod world with a too short mesh chunk length hangs loading - Renovation 2026 DRAGONISLAND.ZEN

オープン
#505 コメント 0 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

まだ誰も着手していません。

評価

難易度
5/5
見積もり時間
1週間以上
初心者へのやさしさ
35/100
issue の種類
バグ
明瞭さ
おおむね明確
活発さ
活発
技術スタック
csharp, unity
領域
backend, game-dev

調査の方向性

The workaround was in ZenWorldArchiveFixDomain.TryFixMeshChunkLengths and ResourceCacheService.LoadWorld; the proposed upstream changes are in ZenKit's World.cc, Mesh.cc, and proto::read_chunked. Start by testing the patched ZenKit with the real DRAGONISLAND.ZEN, then check other mods and rebuild libzenkitcapi for Windows, Quest, and Pico. Done means the world loads in Gothic-Unity and the results are ready for the ZenKit issue and pull request.

索引モデルが issue の本文から書いたものです。

説明

bug

Problem

Renovation 2026 (Gothic 2 mod) world DRAGONISLAND.ZEN makes ZenKit hang (no exception) - the game freezes in PreCaching. The original engine loads it fine.

Cause (ZenKit)

The world mesh's material chunk (0xB020) stores a too small length: 75890 instead of 76067 bytes, its last material is left out. The engine reads the materials one by one and doesn't notice. ZenKit:

  1. Hangs in World.cc: it hops over the mesh chunks by their stated sizes (do { id; seek(len) } while (id != 0xB060)) to read the BSP tree first. With the wrong size it lands inside the last material, walks garbage sizes to EOF, and loops there forever (seeking past the end of an in-memory stream is a no-op, read_ushort() at EOF returns 0, so 0xB060 is never found).
  2. Loses data: proto::read_chunked seeks back to the stated chunk end after the material handler correctly read past it, and reads the rest of the last material as further chunks (vertices/polygons lost or bad_alloc).

Workaround we had (removed again, branch split/23-world)

ZenWorldArchiveFixDomain.TryFixMeshChunkLengths: for every world while a mod was active, ResourceCacheService.LoadWorld copied the world file, walked the ASCII [MeshAndBsp entry's chunk chain (0xB0../0xC0..) and, when a chunk wasn't followed by a known chunk, set its length up to the next spot (within 1 MB) from which the chain continues up to 0xC0FF. Only ASCII archives (binary ones store the entry differently).

Cost to remember for any future workaround: Vfs.Find(name).Buffer.Bytes copies the whole .zen into a managed byte[] (tens of MB for a big world) on every world load while a mod is active - even when nothing needs fixing; the copy was discarded and ZenKit read the VFS again → GC spike on Quest. If a workaround comes back, limit it to worlds shipped by the mod itself and avoid the copy when the chain is fine.

Workaround code
        /// <summary>
        /// Chunk length: a chunk of the world mesh whose stored length is wrong - ZenKit jumps
        /// to the wrong spot, reads the data there as the next chunk and hangs (no exception: the game froze in
        /// PreCaching). Renovation 2026's DRAGONISLAND.ZEN: the materials chunk's length leaves out its last material
        /// (75890 instead of 76067). The engine reads the materials one by one and doesn't notice.
        /// The chunks are walked; a chunk not followed by a known chunk gets the length up to the next spot from which
        /// the chain continues. Returns the fixed copy or null when the chain is fine (or the world isn't an ASCII
        /// archive - the binary ones store the entry differently).
        /// </summary>
        public static byte[] TryFixMeshChunkLengths(byte[] data, out string reason)
        {
            // BIN_SAFE worlds (vanilla) have "[MeshAndBsp" as a string too, but not the raw chunks right after it.
            var tag = IndexOf(data, _asciiFormatTag, 0, Math.Min(data.Length, 256)) < 0
                ? -1
                : IndexOf(data, _meshAndBspTag, 0, Math.Min(data.Length, 4096));
            if (tag < 0)
            {
                reason = "no ASCII [MeshAndBsp] entry";
                return null;
            }

            // The entry's stored size can be wrong as well (DRAGONISLAND.ZEN: short by the same material) - ZenKit
            // doesn't use it, the chunks are walked up to the BSP's end chunk within the file.
            var blob = Array.IndexOf(data, (byte)'\n', tag) + 1;
            var end = data.Length;
            byte[] fixedData = null;
            var fixes = new List<string>();

            var chunk = blob + 8;
            while (chunk + 6 <= end)
            {
                var source = fixedData ?? data;
                var id = BitConverter.ToUInt16(source, chunk);
                if (id == _bspEndChunk)
                    break;

                var length = BitConverter.ToInt32(source, chunk + 2);
                var next = chunk + 6L + length;
                if (length >= 0 && next + 6 <= end &&
                    _meshAndBspChunks.Contains(BitConverter.ToUInt16(source, (int)next)))
                {
                    chunk = (int)next;
                    continue;
                }

                var realNext = FindChunkChain(source, (int)Math.Min(end, chunk + 6L + Math.Max(0, length)), end);
                if (realNext < 0)
                {
                    reason = $"chunk 0x{id:X4} at {chunk}: stored length {length} leads nowhere, no chunk found " +
                             "behind it";
                    return null;
                }

                fixedData ??= (byte[])data.Clone();
                var realLength = realNext - chunk - 6;
                BitConverter.GetBytes(realLength).CopyTo(fixedData, chunk + 2);
                fixes.Add($"chunk 0x{id:X4} length {length} -> {realLength}");
                chunk = realNext;
            }

            reason = fixedData == null ? "chunk lengths are fine" : string.Join(", ", fixes);
            return fixedData;
        }
        private static bool IsChunkChain(byte[] data, int position, int end)
        {
            if (position < 0 || position + 6 > end ||
                !_meshAndBspChunks.Contains(BitConverter.ToUInt16(data, position)))
                return false;
            if (BitConverter.ToUInt16(data, position) == _bspEndChunk)
                return true;

            // long: a garbage length mustn't overflow into a valid looking position.
            var next = position + 6L + BitConverter.ToInt32(data, position + 2);
            return next >= position + 6 && next + 6 <= end &&
                   _meshAndBspChunks.Contains(BitConverter.ToUInt16(data, (int)next));
        }
        private static int FindChunkChain(byte[] data, int start, int end)
        {
            var searchEnd = Math.Min(end - 6, start + _maxChunkLengthError);
            for (var i = start; i <= searchEnd; i++)
            {
                if (data[i + 1] is 0xB0 or 0xC0 && IsChunkChain(data, i, end))
                    return i;
            }
            return -1;
        }

Proposed upstream fix (prepared locally in ZenKit, not submitted yet)

  • World.cc: read the world mesh sequentially (mesh → BSP → triangulate) like the engine, the chunk hopping loop is gone.
  • Mesh.cc (MATERIAL chunk): if the materials alone run past the stated chunk end, the size is wrong - continue after the materials (+ G2 alpha_test byte) and log an error. Not a global "never seek back": ZenKit's own Mesh::save doesn't write the G2 alpha_test byte, its round trip relies on seeking back.
  • proto::read_chunked: a chunk whose stated end lies past the input ends the loop with an error instead of hanging; fixes the underflowing overflow log value.
  • Tests: read_chunked edge cases and a world round trip with a shortened material chunk (fail on current ZenKit, pass with the fix). All 23 vanilla G1/G2/Addon worlds load with identical counts before and after.
  • Not covered: another mesh/BSP chunk with a wrong size (garbage, but no hang); a material chunk short by only the 1-byte alpha_test flag.

Suggested commit: fix(World): read the world mesh sequentially and stop on chunks reaching past the input

Side notes for upstream: Mesh::save doesn't write the G2 alpha_test byte; BspTree::load crashes on an empty BSP tree.

Before we go to the ZenKit author

  • Test the patched ZenKit with the real DRAGONISLAND.ZEN (Renovation 2026) - not done yet, the file wasn't available
  • Check other mods for other wrong chunk sizes (not covered by the fix)
  • Rebuild libzenkitcapi (Windows, Quest, Pico) with the fix and load the world in Gothic-Unity
  • Open the ZenKit issue + PR with the results (GothicKit/ZenKit)
主要言語
C#
スター
33
フォーク
1
平均マージ
2日 4時間
マージ済み PR(30日)
76

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

Gothic-Unity-Project/Gothic-Unity のほかの issue

Gothic-Unity-Project/Gothic-Unity の issue をすべて見る

似ている issue

C# の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。