mo-vidia — Optimized NVIDIA Open GPU Kernel Modules

mo-vidia provides a high-performance kernel optimization patchset and automated Fedora / RPM Fusion RPM packaging pipeline for NVIDIA/open-gpu-kernel-modules.

Kernel Optimizations Included

1. VRAM Physical Fragmentation & Premature OOM Prevention (video_mem.c, mem_mgr_gm107.c):
   • Non-Contiguous-First PMA Allocation: Allocates non-contiguous physical VRAM frames immediately when NVOS32_ATTR_PHYSICALITY_ALLOW_NONCONTIGUOUS is set on non-NUMA discrete GPUs, preserving 2 MB contiguous physical blocks and avoiding failed-contiguous + Copy Engine scrubber stalls.
   • 2 MB $\rightarrow$ 64 KB Page Size Fallback: Automatically falls back from RM_PAGE_SIZE_2M to RM_PAGE_SIZE_64K when a default-page-size $\ge 2\text{ MB}$ VRAM allocation fails 2 MB PMA allocation due to physical fragmentation.
   • Graceful Hardware Compression (comptag) Fallback: Falls back to an uncompressed PTE kind (NVOS32_ATTR_COMPR_NONE) when kmemsysAllocComprResources_HAL exhausts compression resources under heavy VRAM pressure instead of failing the allocation.
2. Fast VRAM Recycling (PMA_FREE_SKIP_SCRUB) (mem_mgr.c):
   • Skips asynchronous Copy Engine scrubbing on free for non-compressible surfaces on non-CC/non-MIG systems while preserving full hardware comptag scrubbing for compressible surfaces.
3. 64 KB Large & 2 MB Huge System Memory Pages & Pools on Discrete GPUs (g_mem_mgr_nvoc.c, mem_mgr.c, system_mem.c, virt_mem_allocator_gm107.c):
   • Enables bSysmemPageSizeDefaultAllowLargePages = NV_TRUE and bAllowSysmemHugePages = NV_TRUE across all discrete GPU architectures (TU102..AD107+).
   • Selects 2 MB (order-9) and 64 KB (order-4) compound pages for aligned system memory allocations, activating the high-order sysmem_page_pools and reducing PCIe GMMU page table entries by $16\times$–$512\times$.
4. Zero-IPI Cache-Type-Aware (WB vs UC/WC) Page Pools (nv-vm.c):
   • Retains NV_MEMORY_WRITECOMBINE / NV_MEMORY_UNCACHED PAT state inside dedicated uc_clean_list and uc_dirty_list pools in nv_page_pool_t, eliminating synchronous cross-core CPU IPI TLB/cache shootdowns (cpa_flush via set_memory_array_uc / set_memory_array_wb) when recycling staging buffers.
   • Keeps sysmem_page_pools active when NVreg_InitializeSystemMemoryAllocations=0.
5. Default PCIe & BAR1 Performance Parameters (nv-reg.h):
   • NVreg_EnableResizableBar = 1 (enables Static BAR1 full-VRAM mapping when Resizable BAR is supported)
   • NVreg_EnablePCIERelaxedOrderingMode = 1
   • NVreg_EnableStreamMemOPs = 1
   • NVreg_InitializeSystemMemoryAllocations = 0
   • NVreg_PreserveVideoMemoryAllocations = 1 & /var/tmp temporary file path for Fedora SELinux compatibility.

---

Installing RPM Packages on Fedora (Drop-in RPM Fusion Replacement)

The built packages (xorg-x11-drv-nvidia-kmodsrc, akmod-nvidia, and kmod-nvidia) use Epoch: 3 and Release: 100.mo_vidia.*, acting as a drop-in upgrade over RPM Fusion's kernel module packages while preserving full compatibility with RPM Fusion's xorg-x11-drv-nvidia, xorg-x11-drv-nvidia-cuda, and nvidia-settings userspace packages.

Option 1: Install from the Automated DNF Repository (GitHub Pages)

sudo dnf config-manager addrepo --from-repofile=https://heretek-ai.github.io/mo-vidia/mo-vidia.repo
sudo dnf upgrade -y xorg-x11-drv-nvidia-kmodsrc akmod-nvidia
sudo akmods --force

Option 2: Install Directly from Local Build or GitHub Releases

sudo dnf install -y ./dist/rpms/*.x86_64.rpm

---

Building RPMs Locally

# Build xorg-x11-drv-nvidia-kmodsrc, akmod-nvidia, and binary kmod-nvidia-$(uname -r) RPMs:
./scripts/build-rpms.sh --with-binary-kmod

# Or clone a specific upstream NVIDIA/open-gpu-kernel-modules tag, apply patches, and build RPMs:
./scripts/build-rpms.sh --upstream-ref 595 --with-binary-kmod

Built .rpm and .src.rpm packages are written to dist/rpms/.