With FEX's host thunks installed, the client runs with a ThunksDB of GL
and Vulkan (FEX_THUNKCONFIG): its Vulkan and GL calls go to the runtime's
own Turnip and Mesa with the session's settings for them, and it gets the
arm64 client's Chromium flags (-cef-force-gpu, ANGLE on Vulkan). Without
the thunks, or with BL_X64_SOFTWARE=1, software rendering as before.
'bannerlator-steam-x64 --gpu' answers for the session.
libblfastpath broke FEX's 32-bit mode (SIGILL at start), so the x86-64
client ran with PROOT_FP_OFF=1 and every one of its syscalls went the
slow ptrace way. In a FEX process it now reads the program FEX is about
to run and stays on for a 64-bit ELF only. bannerlator-steam-x64 drops the
blanket switch (BL_X64_FP_OFF=1 brings it back).
FEX_THUNKS=1 (build-fex.yml 'thunks' input) builds FEX's host thunks: the
arm64 half against the runtime's graphics dev packages, the x86 and i686
guest halves with clang against a sysroot of Arch Linux x86_64 and
multilib packages unpacked by pacman (not run). With them an x86 program's
Vulkan, GL, X11, Wayland and DRM calls reach the runtime's own libraries -
Turnip and Mesa on the GPU - instead of llvmpipe.
- DroidDeck direct: a compatibility tool the x86 client gets in its own
compatibilitytools.d, running a native Linux game with Valve's scout
library runtime under FEX instead of the Steam Linux Runtime container
(pressure-vessel, which needs bwrap). Left 4 Dead 2 (BL_X64_DIRECT_APPS)
is mapped to it in the x86 config while that client is not running.
- The session shim and the controller reader are also built for i386 and
preloaded through glibc's $LIB, so 32-bit games get them too.
- The runtime's bin/ on the client's PATH (steam-runtime-launch-client).
- No -steamdeck -steamos3 for the x86 client yet: its SteamOS audio
manager aborts on the runtime's PulseAudio 13 (pa_operation_get_state).
A Steam setting (Client > Steam client: ARM64, the default, or x86-64
under FEX, experimental) sent as BL_STEAM_CLIENT_ARCH. With x86_64 the
session starts Valve's x86-64 client through bannerlator-steam-x64 in its
own home instead of steamrtarm64/steam, so the Play button gets every fix
the session has (system bus and NetworkManager stand-in, gamescope, Deck
and Big Picture UI, stale locks, branch sync, restart loop, stop watcher,
logs, core pinning); what is only for the arm64 client (ARM64 Proton and
its mapping, redist seeding, Decky, the steamclient.so probes) is skipped.
bannerlator-steam-x64 gains --prepare and --exec for that; its own library
replaces the shared steamapps link.
The client ran on FEX's defaults (half-barrier TSO, full-precision x87).
Use the app's default game preset, PERFORMANCE_TSO (WinNative's), as the
FEX_* variables both FEX builds read; values already set win.
From an audit of bannerlator-session against bannerlator-steam-x64:
- a system bus and the NetworkManager stand-in when none is up, stopped
with the client: the client showed no connection on the desktop
- stale .crash/steam.pid/htmlcache Singleton* cleared before a start
- package/beta matched to -clientbeta before every start (rc=42 loop)
- BL_CLIENT_CPUS re-pinning of the client and its web helper every 5 s
- pathcache.c and perms.c also recognise steamrt64 (the path cache saved
most of the arm64 client's start-up; the write override covers its
self-update)
The session's pads exist for a guest program only through libfakeinput
(the FAKE_EVDEV_* rings), which was built for arm64 alone, so the x86
client saw no controller. Build it natively for x86-64 too, stage it, and
preload it through the FEX rootfs next to the session shim.
Hiding it from the whole x86 web helper got it past tier0 (message loop,
BrowserCreated) but then Chromium trapped on the status descriptor it
opened itself, now a memfd copy - which is the other reason the shim
leaves the web helper alone. With BL_WEBHELPER_HIDE_TRACER=1 only opens
called from libtier0_s.so get the copy; Chromium's get procfs.
Traced per process through FEXServer: every x86 web helper read
/proc/self/status (the string sits in libtier0_s.so) and died of SIGTRAP
right after. proot is a tracer, so tier0 believes a debugger is attached
and breaks into it with int3 on the first failed assertion. The arm64
build only issues a stray svc there, which is why the session shim leaves
the arm64 web helper reading the real TracerPid. BL_WEBHELPER_HIDE_TRACER=1,
set by bannerlator-steam-x64, serves the x86 one TracerPid 0 as well.
A diagnostic: FEX's file log is opened without O_APPEND, so each process
overwrites the last. Through a foreground FEXServer every line carries
the pid and tid it came from.
On X11 Chromium sets up GBM on the render node Xwayland hands out over
DRI3 even with the GPU disabled, loops on KMS dumb-buffer ioctls against
the KGSL device, and never creates the SteamUI browser; the client
restarted it every 10 s. Without a GPU, Chromium's Wayland backend draws
to shared-memory buffers.
Hiding /dev/dri was not enough: Xwayland hands the render node out over
DRI3, and llvmpipe then loops on KMS dumb-buffer ioctls against the KGSL
device. LIBGL_DRI3_DISABLE=1 presents through plain X.
x86 Mesa and the Vulkan loader still enumerated /dev/dri, found the
session's KGSL node and looped on it (KMS dumb-buffer ioctls), and the
webhelper never brought a browser up; the client restarted it every
10 s. FEX resolves paths in the rootfs first, so an empty /dev/dri there
leaves llvmpipe and lavapipe.
The desktop exports zink over the arm64 Turnip ICD (GALLIUM_DRIVER,
MESA_LOADER_DRIVER_OVERRIDE, VK_ICD_FILENAMES); x86 Mesa found no driver
and the client died on glXChooseVisual. Clear them and use llvmpipe.
LD_LIBRARY_PATH beats the system folders regardless of order, so the
platform's own older libc.so.6 shadowed the rootfs's (GLIBC_2.38 not
found). Keep only the libraries the rootfs lacks.
Valve's platform runtime files carry each library under its full version
only; pressure-vessel makes the soname links when it builds a container.
Link them into a folder of our own and run the rootfs's ldconfig -n there,
so libgtk-x11-2.0.so.0 and the rest resolve.
steamui.so needs GTK 2 (libgtk-x11-2.0.so.0), which the SteamRT container
provides and FEX's Ubuntu 24.04 rootfs does not. Append Valve's own
steamrt3c platform library directories to the client's library path,
after the rootfs.
After its self-update the x86 client re-executes itself in the SteamRT
container through pressure-vessel-wrap, which needs bwrap. The FEX rootfs
with the client's steamrt64 libraries is the container here, as proot is
for the arm64 client: set the client's recursion guard. Restart on exit
code 42 the way steam.sh does.
The x86 client stops at start on a tier0 assertion: semget answers ENOSYS,
as Android kernels have no System V IPC (traced with a DEBUG_STRACE FEX).
The session shim that gives the arm64 client its semaphores is now also
built as native x86-64 and preloaded through the FEX rootfs's own
ld.so.preload. SessionFiles stages it and bannerlator-steam-x64 into
installed runtimes.
A workflow_dispatch input passes extra CXXFLAGS (e.g. -DDEBUG_STRACE for a
syscall trace) into the build; such builds stay artifacts and are never
published under a release tag.
The x86 client carries pressure-vessel and will not start when
steam-runtime-check-requirements cannot create a user namespace; no
Android app can, and proot is the container here as it is for the arm64
client, which has no such probe. STEAM_RUNTIME_STEAMRT points it at
Valve's runtime linked file by file, with only that probe replaced.
Through steam.sh, steam-runtime-check-requirements' bwrap probe needs user
namespaces, which proot does not have. Run the SteamRT3 client directly
with STEAM_RUNTIME=1 and its own library folders, after the ~/.steam
links steam.sh would make, as bannerlator-session does for steamrtarm64.
The legacy ubuntu12_32 path in steam.sh refuses to start without user
namespaces (steam-runtime-check-requirements, exit 71). On the beta branch
STEAM_FORCE_CLIENT=steamrt64 makes steam.sh run steamrt64/steam directly,
the x86 twin of the arm64 client the session already runs.
With libblfastpath loaded into FEX, every i386 program died with SIGILL
(FEX's fatal trap) before logging was up, and steam.sh reported the
32-bit libc.so.6 missing. Bisected on device: the session's other two
preloads are harmless; PROOT_FP_OFF=1 lets i386 ld.so run.
Our proot answers openat2 with ENOSYS. FEX opened every rootfs path with
openat2(RESOLVE_IN_ROOT) and only fell back on EXDEV, so in a session
every x86 file lookup missed the rootfs and hit the arm64 guest: bash
under FEX could not find libtinfo. GetEmulatedFDPath already follows
absolute symlinks inside the rootfs, so a plain openat relative to the
rootfs fd is the same lookup. The rootfs also gets an empty
ld.so.preload, so the guest's arm64 preloads are not tried.
FEX processes take the rootfs from FEXServer. The version probe started
one before FEX_ROOTFS was set, and later clients inherited its empty
rootfs. Still not enough on its own on the device: under investigation.
The arm64 Linux client never runs VAC: TF2 and L4D2 are refused by VAC
servers and no client log mentions VAC, because VAC modules are x86 code.
bannerlator-steam-x64 installs Valve's x86-64 client, FEX and FEX's
Ubuntu 24.04 x86 rootfs, and runs the client unchanged under FEX in a
home of its own, sharing only steamapps with the arm64 client. The arm64
client stays the default; the desktop menu gets an experimental entry.
The x86-64 Steam client experiment needs FEX inside the guest, and neither
Arch Linux ARM nor FEX publishes a Linux arm64 binary. Built in the same
Arch Linux ARM container as gamescope, without thunks, tuned for generic
armv8-a, with the runtime's own unsquashfs alongside to unpack FEX's x86
rootfs image.
Left 4 Dead 2 logs "Connecting to", retries, and refuses a VAC server with
a "Connection Blocked: Please remove plug-ins" dialog that never reaches
the console. Counting the attempt as reaching a server would have reported
an insecure game as connected. Only "Connected to" counts now; an attempt
alone is reported as not connected, naming the dialog to look for.
The client reads config.vdf once, as it starts. A title installed during a
session got its per-app ARM64 Proton entry from the compat beat, but the
running client kept its own pick: native for a Linux build, which then sat
on "Downloading content (0%)" waiting for Steam Linux Runtime 2.0 and FEX
it can never run here (Left 4 Dead 2), or proton-stable for a Windows one.
The beat now runs the compat pass with --live, which notes titles it maps
while the client runs. Once each has finished installing and no game is
running, the session restarts the client through the existing restart-once
marker. Once per title and three times per session at most. The mapping is
re-applied before the client starts again, since it writes config.vdf back
from memory as it exits.
A Source-engine game says whether it started VAC-secure only in its own
console, which reaches a file only with -condebug. With the new Setup
toggle on (off by default), the launch wrapper adds -condebug to Source
and Source 2 titles and notes each launch; at the end of the session each
console.log written during it is copied, scrubbed, into vac/ and vac.txt
quotes the game's secure/insecure lines, any VAC lines in the client's
logs and which client builds are on disk. No connection to a server is
reported as inconclusive rather than guessed.