RTX 5070 Ti: GPU VA space exhaustion (dmaAllocMapping can't alloc VA space) leads to Xid 31 and gnome-shell session crash on 610.43.03

[SmookeyDev]

Summary

On an RTX 5070 Ti (Blackwell) running GNOME/Wayland, the driver progressively loses GPU VA space for DMA/framebuffer mappings until the desktop session dies. The failure arrives in waves every few days: dmaAllocMapping_GM107: can't alloc VA space for mapping floods the log, followed by NVRM: ... Out of memory ... mapping_reuse.c:273 asserts, Xid 31 MMU faults in gnome-shell, and [nvidia-drm] Failed to map NvKmsKapiMemory. gnome-shell then SIGABRTs, taking down mutter, Xwayland and GPU-accelerated Electron apps (Discord, VS Code, Chrome) with it. Logging out/in (or rebooting) resets the condition, which strongly suggests a VA-space leak in mapping reuse rather than actual VRAM exhaustion (only ~1.3/16 GiB VRAM in use at idle).

System

• GPU: NVIDIA GeForce RTX 5070 Ti (16 GiB), open kernel modules
• Driver: 610.43.03 (RPM Fusion, Fedora 44)
• Kernel: 7.1.4-200.fc44.x86_64
• Session: GNOME on Wayland, nvidia-drm.modeset=Y, nvidia-drm.fbdev=Y
• Full bug report: https://gist.github.com/SmookeyDev/100931ce97fb50a7b68f95f902ad8cef

Frequency (from journal, since 2026-07-01)

• 390× NVRM: dmaAllocMapping_GM107: can't alloc VA space for mapping.
• 122× NVRM: nvAssertOkFailedNoLog: Assertion failed: Out of memory [NV_ERR_NO_MEMORY] (0x00000051) ... pReuseMappingDb->pMapCb ... @ mapping_reuse.c:273
• 51× Xid 31 MMU faults, all with pid=<gnome-shell>, e.g.:
  Xid (PCI:0000:08:00): 31, pid=4391, name=gnome-shell, channel 0x00000004, intr 00000000. MMU Fault: ENGINE GR_HOST0 HUBCLIENT_ESC0 faulted @ 0x0_04040000. Fault is of type FAULT_INFO_TYPE_UNSUPPORTED_KIND ACCESS_TYPE_VIRT_READ
• 1× NVRM: krcWatchdog_IMPL: RC watchdog: GPU is probably locked! (2026-07-30)
• Each wave ends with: [drm:__nv_drm_gem_nvkms_map [nvidia_drm]] *ERROR* [nvidia-drm] [GPU ID 0x00000800] Failed to map NvKmsKapiMemory 0x000000003b3877f3 followed by gnome-shell/mutter/Xwayland SIGABRT

Example timeline of a full session crash (2026-08-04)

01:16:18  NVRM: dmaAllocMapping_GM107: can't alloc VA space for mapping.  (repeated)
01:16:18  resource: resource sanity check: requesting [mem 0xdfdc0000-0xe00dffff], which spans more than 0000:08:00.0 [mem 0xd0000000-0xdfffffff 64bit pref]
01:16:18  caller __nv_drm_gem_nvkms_map+0xc3/0x130 [nvidia_drm] mapping multiple BARs
01:16:18  [nvidia-drm] *ERROR* Failed to map NvKmsKapiMemory 0x000000003b3877f3
01:16:19  NVRM: Xid (PCI:0000:08:00): 31, pid=4391, name=gnome-shell, ... MMU Fault ...
01:16:24  NVRM: nvAssertOkFailedNoLog: ... Out of memory ... @ mapping_reuse.c:273  (repeated)
01:16:31  systemd-coredump: Process 4391 (gnome-shell) dumped core (SIGABRT)
01:16:36  systemd-coredump: mutter-x11-frames SIGABRT
01:16:37  systemd-coredump: Xwayland SIGABRT
01:17:06  systemd-coredump: Discord SIGTRAP

Notes

• The "Out of memory" here is not VRAM pressure — it is VA space for mappings (pReuseMappingDb), which points to mappings not being freed/reused correctly under a long-running compositor + many GPU-accelerated clients.
• Crash waves correlate with long uptimes (2+ days); a fresh login behaves normally.
• Happy to test a beta/hotfix driver or collect additional traces on request.

[SorenHJohansen]

Adding a reproduction from a different, non-GNOME stack that shows the same
failure can trigger quickly from a Chromium-browser GPU process, not just a
long-running compositor.

System

• GPU: NVIDIA GeForce RTX 3080 (GA102, 10 GiB), open kernel modules
• Driver: 610.43.03
• Distro: Arch Linux (Omarchy), kernel 7.1.4-arch1-1
• Session: Hyprland on Wayland, nvidia-drm modeset
• Browser: Vivaldi 8.1.4087.61 (Chromium 149+, flatpak), launched with
  --ozone-platform=wayland

Repro

1. Open https://store.steampowered.com/sale/steamframe (heavy page: many
   animated banners / textures)
2. Scroll

The Vivaldi GPU process dies within seconds; VRAM is only ~0.7/10 GiB used at
the time. The exact signatures from issue #1274:

NVRM: nvAssertOkFailedNoLog: Assertion failed: Out of memory NV_ERR_NO_MEMORY
(0x00000051) returned from pReuseMappingDb->pMapCb(... _reusemappingdbAddMappingCallback)
@ mapping_reuse.c:273
NVRM: nvAssertOkFailedNoLog: Assertion failed: Out of memory NV_ERR_NO_MEMORY
(0x00000051) returned from reusemappingdbMap(...) @ kern_bus_gm107.c:3150
[drm:__nv_drm_gem_nvkms_map nvidia_drm] ERROR nvidia-drm
Failed to map NvKmsKapiMemory 0x000000002c599937
Vivaldi GPU process terminated abnormally

Notes

• Not VRAM: failure is the pReuseMappingDb VA-space exhaustion you describe.
• Not specific to a long-uptime compositor: happens in seconds here, driven by
  rapid DMA-BUF/compositing mapping churn while scrolling a texture-heavy page.
• Composites fine in software: passing --disable-gpu to the browser stops it.