Library / SDK
strongtz/i915-sriov-dkms avatar
strongtz/i915-sriov-dkms

Six compile probes and two drivers inside one i915 SR-IOV kernel module snapshot

dkms module of Linux i915 driver with SR-IOV support

2,571 stars253 forksCGPL-2.0

At a glance

What is it?
i915-sriov-dkms is a community snapshot of the upstream i915 and xe modules with SR-IOV ported in, packaged for Arch, Debian, Nix and manual installs. It has to be installed on the host and in every guest, and the kernel command line you need changes with the driver you pick.
Who is it for?
Read this module as experimental infrastructure rather than a product: it is a kernel snapshot rebased by a small group of contributors, it is not connected to Intel, and the README says to use it at your own risk. It fits a Proxmox or bare metal host whose Intel GPU has no SR-IOV support upstream and whose guests need their own hardware.
Can I use it commercially?
Yes, with conditions. GPL-2.0 is a copyleft licence: if you distribute software that includes it, you must release that software's source code under the same licence. Running it internally without distributing it does not trigger that obligation.
Is it still maintained?
Yes. The repository last received commits 8 days ago.
What is it written in?
Mainly C, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • strongtz/i915-sriov-dkms2,571 ★ · +41
Star history
Repository2026-09-302026-11-15
strongtz/i915-sriov-dkms2.5k2.6k

41 Stars in the last 46 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

The same package goes on the host and in the guest

The warning at the top says the module has to be installed in both host and guest, and a later note relaxes one dimension of that: the host and the guest VM are not strictly required to use the same module version. What both ends still share is the kernel support matrix, because the 6.18.x to 7.2.x requirement is what decides whether the module loads at all. Packaging is what makes the dual install workable, and the tree carries four answers to it: PKGBUILD covers Arch, debian/ together with i915-sriov-dkms.install covers Debian and Ubuntu, flake.nix with flake.lock covers Nix, and the Makefile is the manual route. The documentation follows the same split, with host guides for Arch, NixOS, Proxmox VE and a manual Debian, Ubuntu and Arch path, then guest guides for Ubuntu, an Ubuntu 25.04 cloud-init VM on Proxmox, and Windows. Only Linux and Windows guests are covered at all, and the Windows guide is pinned to a specific combination: Proxmox 8.3, Windows 11 24H2, and Intel driver builds 32.0.101.6460 and 32.0.101.6259.

Six compile probes decide which kernel APIs the snapshot gets

Spanning many kernel versions without maintaining many branches is handled by conftest.sh, and the Makefile wires the probe step in before any driver source is compiled. A shell block runs at build time, guarded by a test for KBUILD_MODULES and for the absence of the results header, and for each probe it appends a line to conftest/results.h under a banner reading Auto-generated by conftest.sh - do not edit. That header is then force-included into every translation unit through LINUXINCLUDE, so the probe results are available to the sources without a single include directive of their own. The probe list names six upstream symbols:

code
CONFTEST_COMPILE_TESTS := \
  copy_from_user_inatomic_nontemporal \
  drm_exec_for_each_locked_object_no_index \
  drm_fb_helper_alloc_info \
  drm_pagemap_compound_folio \
  pci_resize_resource_4args \
  xe_pmt_telem_read_kernel_device

Four are DRM or PCI kernel interfaces whose signatures and behaviour moved across the releases this snapshot spans, and two are xe telemetry and PMT paths. Each probe compiles a test rather than grepping a header, so an argument count change is caught the same way a missing function is. The conftest directory is registered in clean-files, so a clean build regenerates the header instead of reusing a stale one.

compat/ holds the shims that keep one tree on old kernels

The Makefile builds three object modules, and that list is also the shape of the package: compat/, drivers/gpu/drm/i915/ and drivers/gpu/drm/xe/. The compat directory is the part that absorbs whatever the six probes found, and the include path is widened with a -I pointing at include/, with config.h from that same directory force-included so the shims are in scope before the driver sources are read. One configuration is set outright rather than probed: CONFIG_DRM_GPUSVM := y, carrying a comment saying the repository vendors its own copy of the GPUSVM module. Two version strings are compiled into the module as well, DKMS_MODULE_VERSION and DKMS_MODULE_ORIGIN_KERNEL. The first reads 2026.09.16-sriov, which is the newest release tag with a suffix, and the second reads 7.2.0, the kernel the snapshot was cut from. A third define, DKMS_MODULE_SOURCE_DIR, passes the absolute path of the build tree into the code, which is what lets an out-of-tree module find its own files. The file's last line declares the targets default, clean, modules, load, unload, install and patch as phony, with no recipes attached to any of them.

One kernel command line cannot serve both drivers

The module ships both i915 and xe, and switching between them is done by editing kernel parameters, because each parameter set blacklists the other module. For i915 the line is:

code
intel_iommu=on i915.enable_guc=3 i915.max_vfs=7 module_blacklist=xe

For xe it is:

code
intel_iommu=on xe.max_vfs=7 xe.force_probe=${device_id} module_blacklist=i915

The two share intel_iommu=on and a ceiling of seven virtual functions, and they differ in two ways that matter. i915 needs enable_guc=3 so the firmware is handed off correctly, while xe needs force_probe set to the PCI device ID, which the instructions tell you to read with cat /sys/devices/pci0000:00/0000:00:02.0/device, because xe does not claim that device on its own. The choice is also not free on every machine: the xe module does not support MTL, Meteor Lake, or LNL, Lunar Lake, and those platforms are told to use i915 instead. The result is that the xe command line is wrong on exactly the hardware where people are most inclined to try the newer driver, and a driver that silently does not load looks the same from inside a VM as a driver that loaded fine.

CCS0 stopped being enabled by default on Xe_LP

Version 2026.09.16 changed a default on Xe_LP platforms, which the README defines as TGL, ADL and RPL. From that release CCS0 is no longer enabled by default, and the consequence is documented narrowly: if Windows guest VMs misbehave, append a parameter to the kernel command line, i915.xelp_enable_ccs=1 for the i915 driver and xe.xelp_enable_ccs=1 for xe. No other consumer of that flag is described, and no other change is called out, so an installation that worked on 2026.09.14 can stop working after the upgrade with a fix that has to be chosen per driver. Because the parameter joins the same command line as the earlier ones, a Xe_LP host ends up with a third variant of the line, defined by CPU generation rather than by driver choice. The project's one-line summary still names only the i915 driver while the documentation states the repository provides both i915 and xe, so a search for the project still returns xe specific parameters that the summary never mentions.

The kernel to release mapping overlaps at 6.12 and 6.18

Choosing a version is a four row lookup: the current release 2026.09.16 for 6.18.x through 7.2.x, 2026.03.05.7 for v6.12 through v6.19, 2025.07.22 for v6.8 through v6.12, and a branch named intel-lts-v6.1 for v6.1 through v6.7. The rows overlap, and the overlaps sit on versions people actually run. A host on 6.18 or 6.19 matches both the first and second rows, and a host on 6.12 matches both the second and third, with no stated rule for which one wins. The dates behind the names do not line up either: the release named 2026.03.05.7 was published on 29 August 2026, while the two newer names 2026.09.14 and 2026.09.16 were published on 14 and 16 September 2026. The README says older branches will no longer be maintained and recommends upgrading to a supported kernel. The credits explain how the supported list was assembled: one contributor rebased onto linux-intel-lts for v5.15 and v6.1 and covered v6.1 through v6.4, another rebased onto the mainline-tracking linux/v6.12 branch and covered v6.8 through v6.13, and a third did a large refactor of the repository covering v6.2, v6.5 and v6.8.

Never pass the PF to a VM, but a container can have the PF render node

Virtual functions are created by writing a count into sysfs, and the count is the same seven that appears as max_vfs in both parameter sets:

code
echo 7 > /sys/devices/pci0000:00/0000:00:02.0/sriov_numvfs

Seven is described as the maximum on Intel UHD Graphics. Which function you hand to a guest is the part with consequences: any VF, 00:02.1 through 00:02.7, can be passed to a VM, while the PF must never be passed to a VM because doing so would crash all other VFs. Containers and LXCs get the opposite instruction, to pass the render node of the PF directly. The same device is written two ways on the page, as 00:02.0 in every sysfs path and as 02:00.0 in the passthrough warning, so anyone matching the PF identifier has to know which form to search for. VFs are also not shared out of the box: a separate guide covers blocking VFs on the host for Arch, Ubuntu and Proxmox VE, which matters as soon as one VM has claimed a function.

Recovery means editing the boot command line from the menu

If the module causes a hang or a black screen, the documented escape is the bootloader rather than a package manager. Under GRUB2 you select the kernel at the boot menu, press e, find the line starting with linux and append module_blacklist=i915,xe after a space, then boot with F10 or Ctrl+X. Under systemd-boot the edit key is also e, the same string is appended after a space to the existing line, and Enter boots. Both instructions name both modules, so one blacklist string covers whichever driver loaded. Once the system is up, the next step is to inspect the kernel logs, and if the module is still unstable, to follow the uninstallation steps in the distribution's own installation guide. The Secure Boot note is the other side of loading an out-of-tree module: it requires Secure Boot to be disabled, and anyone who needs Secure Boot has to use a signed kernel and follow the signing guide, which covers Ubuntu, Proxmox VE and Debian-based distributions. Signing is therefore something the installer does rather than something the release hands over, and it applies to the host and every guest copy.

Editorial conclusion

Read this module as experimental infrastructure rather than a product: it is a kernel snapshot rebased by a small group of contributors, it is not connected to Intel, and the README says to use it at your own risk. It fits a Proxmox or bare metal host whose Intel GPU has no SR-IOV support upstream and whose guests need their own hardware. It does not fit a machine where Secure Boot has to stay on, or where nobody can reboot into the bootloader to blacklist a module after a black screen. Before installing, check four things: which kernel you are on against the four row release table, because the rows overlap at 6.12 and 6.18 with no tie break given; whether your CPU generation is one the driver you intend to use actually supports, since xe excludes MTL and LNL; whether the host and guest copies can differ in version, which the notes permit but never explain; and whether you are on 2026.09.16, where CCS0 is no longer enabled by default on Xe_LP and Windows guests may need an extra parameter. The code is GPL-2.0.

Frequently asked questions

What is the Nvidia DKMS package?

Nothing in this repository concerns Nvidia: it builds Intel's i915 and xe kernel modules with SR-IOV support ported in. The DKMS side is visible in the tree, with dkms.conf at the root, a PKGBUILD for Arch, a debian/ directory for Debian-based systems, flake.nix for Nix, and a Makefile whose module version reads 2026.09.16-sriov.

What is DKMS in Linux?

The acronym is never expanded anywhere in this repository, so its definition is not available here. What the tree does show is the packaging side concretely: dkms.conf at the root, distribution packaging for Arch, Debian-based systems and Nix, a Makefile written for an out-of-tree kernel module build, and an install target alongside a conftest.sh that probes kernel symbols at build time.

Where are DKMs modules stored?

The repository does not record where built modules end up on a system; it only shows the inputs. What it does record is a settings file named i915-set-sriov-numvfs.conf in the root, and the sysfs path where the number of virtual functions is written, /sys/devices/pci0000:00/0000:00:02.0/sriov_numvfs.

Official sources

  1. Issues
  2. License: GPL-2.0
  3. README
  4. Releases
  5. strongtz/i915-sriov-dkms on GitHub
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/strongtz-i915-sriov-dkms.svg)](https://hysenlabs.com/projects/strongtz-i915-sriov-dkms)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/strongtz-i915-sriov-dkms.svg)](https://hysenlabs.com/projects/strongtz-i915-sriov-dkms)

Was this useful?