NVIDIA Kernel 470.256.02 Linux Kernel 6.12.43+deb13-amd64 DKMS Fixes

470.256.02_DKMS_FIXES_6.12.43.md

Patching NVIDIA 470.256.02 for Linux Kernel 6.12 -> 6.13 Transition (Debian 13)

Assuming you have the appropriate patches and had DKMS working...

0001-Fix-conftest-to-ignore-implicit-function-declaration.patch
0002-Fix-conftest-to-use-a-short-wchar_t.patch
0003-Fix-conftest-to-use-nv_drm_gem_vmap-which-has-the-se.patch
nvidia-470xx-fix-gcc-15.patch
kernel-6.10.patch
kernel-6.12.patch

Target Environment

• OS: Linux (Debian 13 / Trixie / Sid)
• Kernel: 6.12.x / 6.13.x
• Compiler: GCC 14.x
• Driver: NVIDIA 470.256.02 (Legacy)

---

Identifiable Errors

Users likely encountered one or more of the following blockers:

1. Missing Symbol Table for RDMA: make: *** [/usr/src/ofa_kernel/Module.symvers] Error 1
2. Symbol Pollution / Duplicate Exports: ERROR: modpost: net/vmw_vsock/vsock: 'vsock_addr_validate' exported twice.
3. Compiler Mismatch: The compiler used to compile the kernel (gcc 14) does not match the current compiler.

---

The "In-House" Success Workflow

1. Align the Toolchain

Ensure the kernel and the driver are using the same GCC version.

• Edit /etc/dkms/framework.conf or the driver's dkms.conf:

  export CC="gcc-14"

2. Prune Unused Enterprise Modules

The nvidia-peermem module (RDMA/InfiniBand) is the primary source of ofa_kernel errors.

• Remove the directory: rm -rf /usr/src/nvidia-470.256.02/nvidia-peermem
• Edit dkms.conf: Remove or comment out all entries for BUILT_MODULE_NAME[4] and DEST_MODULE_LOCATION[4] (or those which mention peermem).

3. Sanitize the Build Logic

Prevent the NVIDIA conftest.sh from searching for external RDMA headers.

• Edit /usr/src/nvidia-470.256.02/conftest.sh: Locate and comment out lines setting MLNX_OFED_KERNEL_DIR or searching for /usr/src/ofa_kernel.

4. Fix Symbol Pollution (The modpost Hammer)

To prevent modpost from scanning the entire kernel tree and reporting "exported twice" for networking/virt symbols:

• Update the MAKE command in dkms.conf:

  MAKE[0]="'make' -j`nproc` NV_EXCLUDE_BUILD_MODULES='' KERNEL_UNAME=${kernelver} IGNORE_CC_MISMATCH='1' modules"

5. Execute Clean Build

Flush the DKMS tree to remove stale artifacts and trigger the build:

sudo dkms remove nvidia/470.256.02 --all
sudo CC=gcc-14 dkms install nvidia/470.256.02

---

Technical Insight: Why This Works

The 470 series is legacy. By removing the nvidia-peermem directory and its associated conftest logic, you remove the dependency on ofa_kernel. Using KBUILD_MODPOST_WARN=1 allows the build to complete even if the Debian header metadata contains duplicate symbol definitions, which is common in the transition to Kernel 6.13.