How the Smooth* family keeps Mesa, firmware, and GPU drivers meaningfully newer than Debian stable ships. Applies primarily to SmoothHTPC and SmoothDesktop. SmoothNAS stays headless and does not pull the desktop Mesa stack, but it does install AMD GPU firmware and small ROCm diagnostics for vLLM-style inference plugins.
Debian stable freezes Mesa, linux-firmware, and sometimes LLVM at release time and then only backports security fixes for ~2 years. That's a feature for servers and a problem for:
- HTPC — VA-API / VDPAU / Vulkan-video paths for hardware decode regress on stale Mesa. New codec support (AV1 decode on older GPUs, HEVC 10-bit paths) lands in Mesa continuously.
- Desktop — Vulkan 1.3+ features, HDR path, newer GPU families (RDNA3/RDNA4, Intel Arc/Battlemage, NVIDIA Blackwell) need current drivers. Gaming depends on it.
- Both — new GPUs need firmware blobs that Debian's
linux-firmwareoften lags by months.
We override exactly these three pieces and ride Debian for everything else.
All three live in the common apt suite (see APT_REPO.md):
Tracks linux-firmware.git tip from kernel.org. Rebuilt as Debian drops behind on GPU firmware (typically monthly, sometimes more often when new hardware lands).
- Pure firmware blobs — no code build, just package-and-sign
- Conflicts / Replaces Debian's
firmware-*andlinux-firmware - Small, cheap pipeline
Rebuilt from upstream Mesa releases using kisak-mesa's Debian packaging as the upstream. Clone kisak-mesa's packaging tree, bump the MESA_VER, rebuild in CI. Not a fork — a same-packaging rebuild.
Produces the full Debian mesa package graph (~20 binaries):
libgl1-mesa-dri,libegl-mesa0,libglx-mesa0mesa-vulkan-drivers,mesa-va-drivers,mesa-vdpau-driverslibgbm1,libglapi-mesa-devand-dbgsymvariants
Override mechanism: apt pin on Package: libgl1-mesa-* libegl* libglx-mesa* mesa-* libgbm1 libglapi-mesa pointing at o=RakuenSoftware, a=common at priority 1001 (see APT_REPO.md).
Metapackage that pulls the full graphics set + firmware + non-free driver integrations. Both HTPC and desktop depend on this through smooth-workstation:
Depends:
mesa-vulkan-drivers,
mesa-va-drivers,
mesa-vdpau-drivers,
libgl1-mesa-dri,
libegl-mesa0,
libglx-mesa0,
libgbm1,
linux-firmware-smooth,
intel-media-va-driver-non-free | intel-media-va-driver,
vulkan-tools
NVIDIA handled separately — see below.
Current Intel decode on HTPC/desktop assumes Debian's non-free section is enabled, because the preferred iHD VA-API driver lives there.
SmoothDesktop and SmoothHTPC are first-class Steam / Proton / Wine systems, so the graphics story is not complete with amd64 packages alone.
When those flavors enable i386, smooth-workstation also pulls the matching 32-bit graphics runtime alongside the native one:
libgl1-mesa-dri:i386libglx-mesa0:i386libgbm1:i386mesa-vulkan-drivers:i386
That keeps the 32-bit OpenGL / Vulkan userspace aligned with the Rakuen-built Mesa stack rather than silently falling back to Debian's older i386 graphics packages.
Mesa's RADV, RadeonSI, and llvmpipe link against LLVM. Mesa's LLVM version requirement bumps with each major Mesa release. Our posture:
- Today (Debian 13 / trixie): Debian ships
llvm-19. Current Mesa releases through at least mid-2026 target LLVM 18/19. We depend on Debian'slibllvm19— no override needed. - When Mesa outgrows Debian's LLVM: add an
llvmpipeline to thecommonsuite. This is a significant new build (~1h) with complex packaging. We don't do this speculatively; we do it when Mesa demands it.
The "LLVM bump forced on us" event is the single biggest hidden cost in the graphics stack. Budget for it when it arrives; don't preempt.
Separate track, separate pipeline.
- Open kernel modules (for RTX 20+, increasingly preferred) — Debian packages them; we don't need to fork.
- Proprietary userspace (
nvidia-driver-*) — Debian ships vianon-free; we can either:- (a) Rely on the desktop/HTPC bootstrap path, which already enables Debian
non-free, and tell users to installnvidia-driver-*from there. Simpler; one fewer thing we maintain. - (b) Mirror the relevant
nvidia-driver-*packages into ourcommonsuite as a convenience. More maintenance; fewer steps for users.
- (a) Rely on the desktop/HTPC bootstrap path, which already enables Debian
Leaning (a) at v1 — Debian's non-free is well-maintained, and the Smooth* value-add isn't repackaging NVIDIA.
DKMS modules build against the installed SmoothKernel headers automatically once installed. No special glue needed.
intel-media-va-driver-non-free (the iHD driver) comes from Debian's non-free section. Required for current Intel GPU video decode (Arc, Battlemage, 11th-gen+ iGPUs with new codec support). HTPC and desktop enable Debian non-free at bootstrap so smooth-gfx can pull it as the preferred driver, with intel-media-va-driver (older, in main) retained as a fallback.
| Package | Cadence | Trigger |
|---|---|---|
mesa |
Quarterly majors (25.0, 25.1, ...) + point releases | Upstream tag |
linux-firmware-smooth |
~Monthly | Debian lags a firmware blob we need for a supported GPU |
| NVIDIA mirror (if we do it) | Per Debian nvidia-driver-* update |
Debian push |
| LLVM | Case-by-case | Mesa requirement bumps past Debian |
Rebuilds go through the same testing → main promotion as the kernel. A broken Mesa rebuild is user-visible within a session (GPU apps fail to start); needs a rollback plan. See RELEASE_MODEL.md.
- Flatpak runtime Mesa — Flatpak apps carry their own Mesa via the Freedesktop runtime; we don't need to update that. Helps apps in Flatpak but doesn't help native Kodi/system graphics.
- Userspace GL/Vulkan loaders —
libvulkan1,libegl1,libgl1(the vendor-neutral loaders) come from Debian; no override needed. - Wayland compositors —
kwin(on desktop),cage(on HTPC),wlroots— pulled unchanged from Debian. Mesa changes independently of these.
- Stable vs development Mesa channel. Stick to upstream stable branches (25.0.x, 25.1.x point releases).
main-branch Mesa is too churn-y for a shipped product. Re-evaluate only if a specific feature is stable-only-after-six-months. - AMDVLK. We don't ship it. RADV (in Mesa) is the default Vulkan driver for AMD; AMDVLK is a corner case. If real demand appears, package it separately — don't bundle.