A mod archive numbered 65534 or 65535 is placed under a free index, so no real placement is read as a placeholder. carry() records every payload in one list (settle() still writes each once per install chunk) and keeps one archive index, as write() does.
A carried asset whose install chunk no package names fails when it is carried, and is noted, as it was before it was written once; it no longer fails the whole merge when it settles.
A carried asset was written again for every superbundle that carried it: 1.4 GB of 2.2 GB with 96 mods, which left no room for a live apply. Each one is now written once per install chunk, when its first superbundle settles.
Two fixes taken from worthlessnorms' pull request 16. The bug it was
opened for, rewritten copies read back from the mod's folder, was fixed
separately in 6fd3845.
Mod merge (Engine/Vfs/mod_merge_cas_store.cpp, mod_merge_combine.cpp)
- To combine two mods' edits the merge reads each mod's copy from the
mod's own folder, which means turning its placement in the merged
patch back into the mod's archive and offset. It took the first of the
mod's blocks in that patch archive. A mod added while the game runs
has all its archives appended to the patch's archive 1, so a copy in
its second archive was looked for in its first, the read failed, and
the asset kept the highest-priority copy.
- unshift() picks the block the offset falls in. Of two blocks that
start at the same byte the later one is used: the earlier is an empty
archive.
- Reproduced with a cosmetic mod rebuilt as two archives and added to a
running merge: cas_main_bundlereftable "could not be combined (Cannot
read ...\cas_01.cas)" before, "combined 2 edits (101 preset(s))"
after. A launch merge gives every archive an index of its own and was
not affected; 40 installed mods merge exactly as before.
- mod_merge_unshift_tests covers an archive with its own index, three
appended blocks, bytes before the first block, and an empty archive.
Tests (CMakeLists.txt, Extension/Multiplayer/Test/mesh_tests.cpp)
- multiplayer_lobbies did not link: steam_lobbies.cpp calls
contains_bad_words and the target did not link dingosdk_word_filter.
- multiplayer_mesh did not link: four functions the session code calls
had no stub. Two are the ones the pull request adds
(update_developer_hoodie, remote_skater_generation); the developer
board has since added update_developer_board and remote_board_entity.
- Both targets build and their tests pass.
Co-authored-by: worthlessnorms <110003169+worthlessnorms@users.noreply.github.com>
Mod archives (Engine/Vfs/mod_merge*)
- A free archive index no longer has to sit at or below the highest index
the layout already declares. The engine keeps archives in a table keyed
by layer, install chunk and a 16-bit index, with no range to stay
inside; what it cannot survive is a reference to an archive the layout
never declared. Indices are taken lowest first, per package directory.
- So a mod that collides takes an index of its own and its file is hard
linked rather than concatenated, and no two mods share one archive's
4 GB of addressable bytes. Sharing is left for a mod added while the
game runs, which has no declared index to take, and for a directory
with no index left.
- A placement addresses its payload with a 32-bit offset, so both append
paths now refuse to cross 4 GB instead of wrapping round and pointing
the game at some other payload: CasStore::write fails with the archive
that is full, and the live-apply path checks archive 1's size plus the
mod's before appending and asks for a restart, which gives the mod an
archive of its own anyway.
Injection through a hooked LoadLibraryW (Engine/Core/Platform)
- A player could not launch at all: "Child LoadLibraryW entry bytes match
neither Windows' kernel32 file nor this process". Anti-virus and
overlays hook the entry with a jump to a per-process trampoline, so the
bytes differ in the launcher and in the child and neither comparison
could match. The guard already allowed the same hook when both
processes carried it identically, so it refused on an accident of how
the hook was written rather than on intent.
- inline_jump_length recognises the forms a hooking engine writes (rel32
jmp or call, push/ret, jmp [rip], mov rax/jmp rax, short jmp). When one
is present and every byte after it still matches the clean kernel32
file, the launch goes ahead: the body is Windows' own, and the provider
image already matched on timestamp, checksum and size. A wholesale
replacement is still refused, and now says so.
- The failure names the modules loaded into Skate that came from neither
Windows nor the game folder, instead of guessing at "security
software". An accepted hook is logged as a warning with the same list.
Launching (Launcher/launch.cpp)
- Exporting loose scripts is a convenience, and a folder that would not
take it was failing the whole launch: fs::create_directories threw and
nothing caught it. The export is skipped, RESKATE_LOOSE_FILES goes to
0 so the runtime does not look for files that are not there, and Skate
starts. Someone whose install is under Program Files can play; only
editing loose scripts is lost.
Failures worth reading (Launcher/problem.h)
- A hooked LoadLibraryW, a refused write and ERROR_ELEVATION_REQUIRED
each explain themselves now. Windows reports the last two in the user's
own language -- "Zugriff verweigert", "La operación solicitada requiere
elevación" -- so the rules key on the operation names and the error
number, which are never translated.
Saved skaters (Extension/Customization, Extension/Profile)
- Outfits stopped saving and every slot showed the standard skater. The
load hook marks a preset blocked on entry and clears it only on a full
success, so a load refused for a catalog that was not ready yet left
the preset blocked for the session and every later save was refused as
previous_load_rejected. The game asks once per level load and does not
ask again, so one lost race cost a skater until a restart. Logs show
the catalog arriving three seconds after the refused load, at startup
and again on a map switch 17 minutes later.
- The record is remembered in CosmeticRuntime::pending_loads and applied
by retry_pending_cosmetic_loads from the customization pump, which
already waits for refresh_cosmetic_catalog. The preset stays blocked
until a load really succeeds, so nothing is written over the outfit on
disk.
- Those records belong to the loadout manager's array, which can be
replaced in between, so each pointer is checked against the live array
exactly as a save checks one -- manager current, begin/end sane, 0x48
stride, at most 10 records, in range and aligned. A record that no
longer belongs is dropped as retry_record_moved rather than written
through.
A player whose game sat under a folder named U+200E (left-to-right mark,
an invisible directory on their Desktop) could not launch: the status
tile read "No mapping for the Unicode character exists in the target
multi-byte code page". Nothing was wrong with their mods, and the merge
in their log had just succeeded.
fs::path::string() converts to the process's ANSI code page and throws
exactly that when a character has no mapping there. initfs.cpp called it
on the normal path, not in an error handler, so exporting loose files
threw on every launch. Their log stops right where that export begins.
Cyrillic, CJK or accented characters in an install path do the same on a
code page that cannot hold them.
- path_utf8() in Engine/Core/Platform/path_text.h: UTF-8 represents every
path Windows allows, so it never throws.
- Every path that becomes a narrow string uses it -- 35 of them across 13
files. Most are log lines and error messages, which were a second trap:
reporting a problem with a path threw this instead of the problem.
- The ~45 .string() calls on Json values are untouched; those are not
paths.
Nothing changes for an ASCII path, where both encodings agree, and a path
that is not ASCII had no working behaviour to preserve.
Reported with a log by a player on Discord. Moving the game to an ordinary
folder works around it on an unpatched build.
worthlessnorms