VKD3D-proton detects Broken NULL PRT with SMEM and applies per-game Starfield workarounds (2 configs + shader quirks), but the workarounds are incomplete. During terrain/material loading, a GPUVM page fault occurs from the SQC data client (SMEM scalar load) at address 0x0000000000001000, causing a full GPU reset with VRAM loss.
dmesg sequence:
Process Main pid 16080 thread vkd3d_queue pid 16925
Ring gfx_0.0.0 reset failed
GPU reset begin!. Source: 1
MES(1) failed to respond to msg=REMOVE_QUEUE (x5)
failed to halt cp gfx
gfxhub page fault (src_id:0 ring:24 vmid:6 pasid:283)
in page starting at address 0x0000000000001000 from client 10
GCVM_L2_PROTECTION_FAULT_STATUS:0x00601431
Faulty UTCL2 client ID: SQC (data) (0xa)
MORE_FAULTS: 0x1, PERMISSION_FAULTS: 0x3, RW: 0x0
MODE1 reset
VRAM is lost due to GPU reset!
Reproduction: On save load, always during terrain/material loading. Always the same SQC data client fault at 0x1000. Same fault signature reported on RX 9070 XT (mesa#15689, mesa#15708) and RX 7900 XTX (mesa#15737) GitLab for MESA
VKD3D-proton log (relevant lines from steam-1716740.log):
3503.886:info:vkd3d-proton:vkd3d_bindless_state_add_binding: Device supports VK_EXT_descriptor_buffer! (x3)
3503.786:info:vkd3d-proton:vkd3d_instance_apply_application_workarounds: Detected game Starfield.exe, adding 2 configs, removing 0 configs.
3503.786:info:vkd3d-proton:vkd3d_instance_apply_application_workarounds: Detected game Starfield.exe, adding shader quirks for specific shaders.
The game runs until terrain/material loading, then VKD3D-proton never receives a device_lost callback — the GPU hangs so hard the kernel resets it before the error propagates. The last game output before the hang: "TERRAIN: Loaded 72 material slots".
Software information
Starfield v1.16.244, heavily modded (~200 ESM mods, 14 SFSE plugins). SFSE v0.2.21. Custom shaders via SFShaderInjector. Launch via
sfse_loader.exethrough gamescope (Wayland, 2560x1080@200Hz, HDR enabled).System information
Issue
VKD3D-proton detects
Broken NULL PRT with SMEMand applies per-game Starfield workarounds (2 configs + shader quirks), but the workarounds are incomplete. During terrain/material loading, a GPUVM page fault occurs from the SQC data client (SMEM scalar load) at address0x0000000000001000, causing a full GPU reset with VRAM loss.dmesg sequence:
Reproduction: On save load, always during terrain/material loading. Always the same SQC data client fault at
0x1000. Same fault signature reported on RX 9070 XT (mesa#15689, mesa#15708) and RX 7900 XTX (mesa#15737) GitLab for MESAAlready tested with no effect: disabled raytracing, disabled frame gen/upscaling, multiple Proton versions (Experimental, CachyOS native), LTS kernel, deleted vkd3d-proton cache, removed
VKD3D_CONFIG=force_host_cached.VKD3D-proton log (relevant lines from
steam-1716740.log):The game runs until terrain/material loading, then VKD3D-proton never receives a device_lost callback — the GPU hangs so hard the kernel resets it before the error propagates. The last game output before the hang:
"TERRAIN: Loaded 72 material slots".Log files
steam-1716740.log(attached)