Commit 419e4d18 authored by Gaetan Lepage's avatar Gaetan Lepage
Browse files

nixos/hardware/nvidia: eager load nvidia-uvm for open driver 

Fixes contingent CUDA issues with open driver:
https://github.com/NixOS/nixpkgs/issues/334180



Co-authored-by: default avatarSomeone Serge <else@someonex.net>
parent 9fb5d292

nixos/modules/hardware/video/nvidia.nix

+16 −12
Original line number Diff line number Diff line
@@ -343,14 +343,23 @@ in 
            ];



            # Don't add `nvidia-uvm` to `kernelModules`, because we want

            # `nvidia-uvm` be loaded only after `udev` rules for `nvidia` kernel

            # module are applied.

            # `nvidia-uvm` be loaded only after the GPU device is available, i.e. after `udev` rules

            # for `nvidia` kernel module are applied.

            # This matters on Azure GPU instances: https://github.com/NixOS/nixpkgs/pull/267335

            #

            # Instead, we use `softdep` to lazily load `nvidia-uvm` kernel module

            # after `nvidia` kernel module is loaded and `udev` rules are applied.

            extraModprobeConfig = ''

              softdep nvidia post: nvidia-uvm

            '';



            # Exception is the open-source kernel module failing to load nvidia-uvm using softdep

            # for unknown reasons.

            # It affects CUDA: https://github.com/NixOS/nixpkgs/issues/334180

            # Previously nvidia-uvm was explicitly loaded only when xserver was enabled:

            # https://github.com/NixOS/nixpkgs/pull/334340/commits/4548c392862115359e50860bcf658cfa8715bde9

            # We are now loading the module eagerly for all users of the open driver (including headless).

            kernelModules = lib.optionals useOpenModules [ "nvidia_uvm" ];

          };

          systemd.tmpfiles.rules = lib.mkIf config.virtualisation.docker.enableNvidia [

            "L+ /run/nvidia-docker/bin - - - - ${nvidia_x11.bin}/origBin"
@@ -639,16 +648,11 @@ in 
          boot = {

            extraModulePackages = if useOpenModules then [ nvidia_x11.open ] else [ nvidia_x11.bin ];

            # nvidia-uvm is required by CUDA applications.

            kernelModules =

              lib.optionals config.services.xserver.enable [

            kernelModules = lib.optionals config.services.xserver.enable [

              "nvidia"

              "nvidia_modeset"

              "nvidia_drm"

              ]

              # With the open driver, nvidia-uvm does not automatically load as

              # a softdep of the nvidia module, so we explicitly load it for now.

              # See https://github.com/NixOS/nixpkgs/issues/334180

              ++ lib.optionals (config.services.xserver.enable && useOpenModules) [ "nvidia_uvm" ];

            ];



            # If requested enable modesetting via kernel parameters.

            kernelParams =