On 2026. július 1., szerda 19:20:49 közép-európai nyári idő WhoAm I wrote: > Hello Mesa developers. > > I am reporting a bug in RADV that I have been investigating for over 200 > hours. I have tested numerous configurations and believe I have isolated > the issue.
Hi, You are mixing up 3 different problems here: (1) GPU freeze/hang with Metro Exodus on Proton (2) The AMDGPU kernel can't recover your GPU after the hang (3) Segfault with Metro Exodus native (1) and (3) are most likely a RADV bug. Possibly two different bugs between the different versions of the game. Please open an issue here: https://gitlab.freedesktop.org/mesa/mesa/-/work_items Select the Radeon Vulkan bug report template and fill in the details, especially the steps to reproduce. Also please upload a full dmesg log after the freeze/hang. (2) is a kernel bug. With Linux 6.18 and newer, the kernel should recover the GPU and should not freeze your full system. Please open an issue here: https://gitlab.freedesktop.org/drm/amd/-/work_items Also mention the steps to reproduce the issue and upload the full dmesg log after the freeze/hang. Best regards, Timur > System Information: > GPU: AMD Radeon RX 6800 XT (Navi 21) (ASRock Phantom Gaming) > CPU: Intel Xeon E5-2697A v4 > RAM: 32 GB > OS: Gentoo Linux > Kernel: 7.0.11-gentoo (also tested 6.18.6 and 7.1.2-cachyos) > Mesa version: 26.1.2 (also tested 26.1.3) > Desktop: bspwm (X11) > Game: Metro Exodus (native Linux version from Steam) > Game settings: DX12, Ray Tracing Ultra, Quality on "Koshmar" (Nightmare / > Extreme) > What I Tried (and what did NOT help): > > Different Proton versions (7.0, 8.0, 9.0, Experimental, GE-Proton) - all > crashed or froze. > Custom Wine build (wine-staging 11.11 with ntsync) - same crashes. > > Proton with VKD3D-Proton was especially problematic - the game would not > just crash, but the entire system would freeze or the GPU would reset. > During these events, dmesg showed: > "Dumping IP State" > "GPU reset begin" > "GPU reset succeeded" > "VRAM is lost due to GPU reset" > "[drm] device wedged, but no recovery needed" > > After this, the system would become completely unresponsive, sometimes even > the keyboard (Caps Lock indicator) would stop working, requiring a hard > reboot. > Mesa environment variables: > > RADV_DEBUG=nodcc - this stopped the crashes in Proton but caused a ~30% > performance drop. RADV_PERFTEST=rt - no effect. > RADV_DEBUG=nodcc,noics - no effect. > RADV_PERFTEST=no_shader_cache - no effect. > > VKD3D environment variables: > > VKD3D_CONFIG=dxr - no effect. > VKD3D_CONFIG=dxr,no_upload_hvv - no effect. > VKD3D_CONFIG=force_bindless_texel_buffer - no effect. > > Vulkan driver switching: > > RADV (default) - crashes in Proton, segfaults in native. > AMDVLK - also crashes in Proton; not tested in native. > > Kernel boot parameters: > > amdgpu.dcdebugmask=0x10 > amdgpu.dcdebugmask=0x12 > amdgpu.dcdebugmask=0x210 > amdgpu.dcdebugmask=0xA10 > amdgpu.gpu_recovery=0 > amdgpu.gpu_timeout=20000 > amdgpu.runpm=0 > amdgpu.mcbp=0 > amdgpu.ppfeaturemask=0xfffd7fff > > None of these resolved the issue. > > Replaced the SSD (the game was on a failing SSD, then moved to a healthy > HDD) - the issue persisted on the HDD, proving it is not storage-related. > Reflashed the motherboard BIOS - no effect. > Tested on a clean Arch Linux live USB - the issue still occurred. > Tried XFS and ext4 filesystems - no difference. > > Comparison: DX12 vs DX11: > > On Proton (DX12 with Ray Tracing Ultra and "Koshmar" quality settings): the > game crashes and sometimes causes full system freezes. > On Proton (DX11 without ray tracing): the system crashes disappear, but > performance drops by about 20%, and the graphics quality is noticeably > worse. This suggests that the issue is specifically related to how RADV > handles the combination of DirectX 12, Ray Tracing, and high-quality > rendering in this game. > The Breakthrough: > > I decided to test the native Linux version of Metro Exodus (not Proton). The > native version does not crash, but during loading screens (specifically the > scenes with Artyom's narration), the game freezes for 3-4 seconds, then > continues. During these freezes, dmesg shows a segmentation fault in > libvulkan_radeon.so: > Jul 1 19:13:52 gentoo kernel: video[270346]: segfault at 40 ip > 00007f7e39fdd275 sp 00007f7d9a7fd350 error 4 in > libvulkan_radeon.so[1dd275,7f7e39e10000+6b3000] likely on CPU 22 (core 6, > socket 0) Jul 1 19:13:52 gentoo kernel: Code: 14 8a 48 ff c1 c5 fe 7f 00 > c5 fa 7f 40 20 c7 00 95 94 9f 3b 4c 89 40 10 44 89 50 28 c5 fa 7f 48 18 49 > 39 cb 74 1f 4d 8b 04 cc <41> 83 78 40 01 75 a4 45 84 ed 74 9f 48 8b 47 18 > c5 fa 7e 0c c8 eb Jul 1 19:14:01 gentoo kernel: video[296182]: segfault at > 40 ip 00007fca867dd275 sp 00007fc9eb7fd350 error 4 in > libvulkan_radeon.so[1dd275,7fca86610000+6b3000] likely on CPU 8 (core 8, > socket 0) Jul 1 19:14:01 gentoo kernel: Code: 14 8a 48 ff c1 c5 fe 7f 00 > c5 fa 7f 40 20 c7 00 95 94 9f 3b 4c 89 40 10 44 89 50 28 c5 fa 7f 48 18 49 > 39 cb 74 1f 4d 8b 04 cc <41> 83 78 40 01 75 a4 45 84 ed 74 9f 48 8b 47 18 > c5 fa 7e 0c c8 eb > The game recovers from this segfault and continues to run, but the freeze is > reproducible every time a location is loaded. > What This Suggests: > > The native version of Metro Exodus exposes a bug in RADV > (libvulkan_radeon.so) that causes a segmentation fault. The game handles > this fault and continues. However, when the Windows version of the game is > run through Proton/VKD3D, this same bug leads to an unrecoverable crash > (the game freezes and eventually dies) or even a full system freeze with > GPU reset. > Conclusion: > > This appears to be a bug in RADV, specifically in how it handles certain > Vulkan operations in Metro Exodus with DX12 and Ray Tracing. The issue is > reproducible on RX 6800 XT and may affect other RDNA2 users. > If you need more logs, core dumps, or specific system information, I am > happy to provide them. > Note: I am a Russian speaker. This report has been translated using a > translator. Please forgive any grammatical inaccuracies. I hope the > technical content is clear. > Thank you for your work on Mesa.
