Files
DroidDeck/docs/arm-v8.0-steam-client.md
T

6.8 KiB

Steam client on ARMv8.0 CPUs

The crash

Valve's native arm64 Steam client stopped starting on ARMv8.0 CPUs after the April 15 2026 build (client version 1776387948). The process dies with SIGILL right away, the session log shows "$steam_root/steamrtarm64/steam" ... Illegal instruction and steam exited rc=132, and the screen stays black.

The faulting instruction is an ARMv8.1 LSE atomic. The GDB trace in ValveSoftware/steam-for-linux#13288 stops on ldaddal x1, x0, [x0], with swpa and casal next to it. Earlier builds were most likely compiled with -moutline-atomics, which picks LL/SC or LSE at run time. Newer builds inline LSE. Valve has not replied on that issue, so the cause comes from the reporter's disassembly, not from Valve.

Affected cores are the ARMv8.0 ones without LSE: Cortex-A35, A53, A57, A72 and A73. A55, A75 and later have it. On Android that covers the Snapdragon 4xx/6xx parts built on A53/A73 (for example the Snapdragon 685 with Adreno 610 in Droid-Deck/DroidDeck#380), plus the Nintendo Switch, Raspberry Pi 4 and older, and the Jetson Nano outside Android. /proc/cpuinfo shows it: the Features line has no atomics.

LSE is an ARMv8.1 feature. LSE-Emu users report the current client running on a Cortex-A57, so nothing later than LSE seems to be required yet.

What other projects do

  • SildurFX/Switchdeck (Switch, L4T Linux) lets the client update once, then overlays linuxarm64, steamrtarm64, steamui and a steam.cfg from the April build. It works there because the install script controls the whole system. For us a pinned client is replaced by the self updater on its next start, which is why #380 was closed.
  • SildurFX/LSE-Emu is a preload library by the same author, written to run the current arm64 client on the Switch. A SIGILL handler performs each LSE instruction with an LL/SC loop. With LSE_EMU_PATCH=1 the first trap at a site writes an LL/SC stub within branch range and replaces the instruction with a b to it, so the site runs natively after that. It covers the full LSE set plus LDAPR. The author reports Steam, FEX, Proton games and ARMSX2 running on a Cortex-A57, with ARMSX2 close to native ARMv8.0 speed once patched. GPL-3.0, same as this repo.
  • Azkali/proton-arm64-a57 rebuilds Proton 11 arm64 (Wine, DXVK, vkd3d-proton, FEX's WoW64 and ARM64EC DLLs) with -march=armv8-a -mtune=cortex-a57. Stock Proton 11 arm64 also SIGILLs on ARMv8.0 (wineboot rc=132), so the client is not the only problem for games.
  • Some installers fall back to Valve's x86-64 client under FEX on CPUs without LSE. We already have that path (BL_STEAM_CLIENT_ARCH=x86_64, droiddeck-steam-x64), but it is slower and FEX itself has to be built without LSE.
  • FrankNat996/armor notes a separate problem: the arm64 client's ELF segments are 4K aligned, so it SIGSEGVs at once on 16K page kernels.

This change

The session vendors LSE-Emu (tools/linuxfs/lse-emu, upstream commit 8bc9548). CI and tools/build_local.sh build it as usr/local/lib/droiddeck-lse/liblse_emu.so with -march=armv8-a -mno-outline-atomics. The build fails if the library contains LSE instructions itself.

When /proc/cpuinfo reports any core without atomics, SessionFiles puts the library first in the rootfs /etc/ld.so.preload. BwrapSpawner already turns that file into LD_PRELOAD for sandboxed processes. On CPUs with LSE nothing changes: the file is staged but not preloaded, and the library would also stay idle on its own because it checks HWCAP_ATOMICS.

droiddeck-session sees the library in ld.so.preload and exports LSE_EMU=1 and LSE_EMU_PATCH=1, so the Steam client and everything it starts get the emulation. With a session debug folder it also sets LSE_EMU_LOG=$BL_DEBUG_DIR/lse_emu.log, and session.log gets a line starting with == cpu: no LSE atomics.

The app targets SDK 28, so the app domain still has execmod and the in-place patching can make file-backed code pages writable. If the target SDK goes up, mprotect on those pages is refused, every site logs mprotect-failed, and execution stays on the slower trap path.

Tests

tools/tests/test_lse_emu.py builds the library and a program that uses every LSE family (ldadd in all four sizes, ldclr, ldeor, ldset, swp, ldsmax, ldumin, casal, caspal) from four threads, and runs it under qemu-aarch64 -cpu cortex-a57:

  • without the library the program dies with SIGILL
  • trapping alone runs every form and keeps the threaded counters exact
  • patching runs every form and patches every site, single threaded
  • without LSE_EMU=1, and on -cpu neoverse-n1, the library stays out of the way
  • a SIGILL handler the program installs itself still runs for an unrelated udf

Threaded patching segfaults under qemu-user, writing to the stub arena, which is mapped and writable. qemu write-protects pages it has translated code from, and that races when other threads run code on them. It needs checking on real hardware.

Testing on a device

On an ARMv8.0 phone or handheld (no atomics in /proc/cpuinfo):

  1. Install the test APK and start a Steam session.
  2. session.log should have == cpu: no LSE atomics and no steam exited rc=132.
  3. In the session debug folder, lse_emu.log should show start, patched and exit lines. cut -d' ' -f2 lse_emu.log | sort | uniq -c summarises it. unhandled means an instruction the library does not cover, and mprotect-failed or no-arena mean sites left on the trap path.
  4. Compare UI responsiveness against an ARMv8.2 device, and note any lse_emu: unhandled lines in session.log.

Still open

  • Games. Proton 11 arm64 and Valve's FEX build need the same treatment. The preload reaches them when they inherit LD_PRELOAD, but pressure-vessel and static binaries can drop it, and FEX installs its own signal handlers with raw syscalls. An ARMv8.0 Proton build like Azkali's is the cleaner fix.
  • In #380 the session first died with SIGILL in the app's own wl-compositor thread, inside turnip-sdk36/vulkan.ad08XX.so, which is the A8xx driver build loaded on an Adreno 610. Picking the 6xx/7xx Turnip fixed it, and Auto picks that build once the driver release list is loaded. Nobody has checked yet whether our bundled driver builds avoid LSE. The preload cannot help there because the compositor is a bionic thread in the app process.
  • JIT code (anonymous memory) is never patched and always traps.
  • If Valve starts requiring something past LSE, the library prints lse_emu: unhandled pc=... insn=... and the client crashes as before.