Skip to content

Latest commit

 

History

History
159 lines (118 loc) · 9.2 KB

kernel-config.md

File metadata and controls

159 lines (118 loc) · 9.2 KB

Kernel config

The canonical .config files for linux-smoothkernel. One config per Debian architecture, shared by every Smooth* flavor. See KERNEL.md for the one-kernel-line rationale and bumping-kernel.md for how the configs carry forward across kernel bumps.

Custody

The configs live in smoothkernel (colocated with the recipes that consume them), committed as configs/smooth-amd64.config and configs/smooth-arm64.config. They are versioned alongside the patch series for that kernel version.

smoothkernel/
├── configs/
│   ├── smooth-amd64.config          Canonical amd64 config, current kernel line
│   ├── smooth-arm64.config          Canonical arm64 config, current kernel line
│   └── <kernel-version>/            Archived configs from prior kernel bumps
├── patches/
│   ├── cachyos-<kernel-version>/    Vendored base lane (`0001-bore.patch` on pristine kernel.org today)
│   ├── nobara-picks/                Cherry-picked Nobara HID patches
│   └── post-nobara-<kernel-version>/ Extra carry patches applied after Nobara
└── recipes/
    └── build-kernel.sh              Consumes configs/ + ordered patch lanes automatically

Invariants

The following settings are load-bearing across every flavor. Changing one has cross-flavor consequences; don't touch without updating KERNEL.md and filing a PR with rationale.

Setting Value Reason
CONFIG_PREEMPT y Full preemption; BORE makes the server-side cost negligible
CONFIG_HZ 1000 UI/game responsiveness; negligible server impact on modern hardware
CONFIG_SCHED_BORE y BORE scheduler enabled by the vendored base lane; default y
CONFIG_SCHED_EXT m sched-ext available as a module; not default, escape hatch
amd64 microarch baseline x86-64-v2 Inclusivity for HTPC/NAS on ~2009+ hardware
arm64 baseline generic Debian arm64 Broadest viable arm64 support; board-specific boot enablement is outside this config
CONFIG_MODULE_SIG y Modules carry signatures; packaged modules are signed at build time and DKMS modules on-host via smooth-secureboot. See signing.md.
CONFIG_MODULE_SIG_FORCE not set (target: y) Aspirational, not current state — see the note below before "fixing" this.
CONFIG_DEBUG_INFO n Set by STRIP_DEBUG_INFO=1 default in build-kernel.sh to avoid debug package bloat
CONFIG_DEBUG_INFO_BTF n Also disabled by the same stripped-debug profile
CONFIG_SYSTEM_TRUSTED_KEYS build-time injected Rakuen module cert Release builds inject the public cert for packaged-module signing; the checked-in config does not carry secrets or machine-local paths
CONFIG_SYSTEM_REVOCATION_KEYS "" Explicitly managed by our release process rather than inherited from Debian packaging defaults

Note: module-signature enforcement is not on, and cannot simply be switched on

CONFIG_MODULE_SIG_FORCE has never been set in a shipped SmoothKernel config — not in 6.19.12, 7.0.11, or 7.1.8. Earlier revisions of this table listed y as if it were current state. It was not, and turning it on today breaks storage. Measured on a 7.1.8 test VM:

Kernel / mode lockdown unsigned module zfs (DKMS-signed)
Debian, Secure Boot on [integrity] rejected n/a
SmoothKernel, default [none] loads (taint 12289) loads
SmoothKernel, lockdown=integrity [integrity] rejected rejected

Enforcement rejects the DKMS module too, because the per-host MOK key is only trusted once it has been enrolled through UEFI — which requires Secure Boot, which requires a kernel signed by a trusted key. Enabling enforcement before that chain exists makes ZFS (and smoothfs) unloadable on every appliance.

Note also that Debian gets its Secure-Boot-triggered enforcement from CONFIG_LOCK_DOWN_IN_EFI_SECURE_BOOT, which is a Debian downstream patch. That symbol does not exist in a pristine kernel.org tree, so it cannot be enabled from scripts/config; olddefconfig silently drops it. Adding it would mean carrying the patch in post-nobara-<version>/.

The enforcement contract and its required ordering live in signing.md. Do not raise enforcement here in isolation.

Filesystems

Built in as modules (available but not always loaded):

  • ext4, xfs, btrfs — all built in as =m
  • bcachefs — external DKMS module, installed from the bcachefs APT repo when a flavor enables it
  • ZFS — not in-tree (CDDL/GPL); ships as openzfs-zfs-dkms
  • NTFS3 — =m (useful for external drives on desktop/HTPC)
  • f2fs, exfat, FAT — =m
  • Network filesystems: CIFS, NFSv3/v4 — =m

Networking

  • IPv4/IPv6 — built in
  • netfilter + nftables — built in (required for smoothrouter's default-deny posture)
  • tcp_bbr — built in, default congestion control (matches existing build-kernel.sh NET_TUNING)
  • sch_fq — built in, default qdisc
  • VLAN, bridge, bonding, VXLAN — =m
  • WireGuard — built in (required for smoothrouter)

Hardware support

Built in: the handful of drivers needed to boot every supported platform. Everything else as modules.

  • Core: AHCI, NVMe, USB storage
  • GPU: amdgpu, i915, xe, nouveau, radeon — all =m; amd64 also keeps DRM_AMDGPU_USERPTR, HSA_AMD, and HSA_AMD_SVM enabled so headless SmoothNAS hosts expose the ROCm /dev/kfd path needed by vLLM-style inference containers.
  • Network: common NIC drivers (e1000e, igc, ixgbe, r8169, iwlwifi, mt76) — =m
  • USB HID, input — =y for boot-time usability

What we drop

The existing build-kernel.sh has an APPLIANCE_TRIM profile, but it is scoped to hardware families no Smooth* flavor is expected to rely on. It must not disable HTPC / Desktop basics like DRM, audio, media, wifi, or controller input. The trim list stays narrow:

  • Mainframe / s390 / IBM Power drivers — n
  • Niche industrial buses (CAN, I2C multiplexers for industrial) — n
  • Exotic network devices (InfiniBand in some configs) — n, revisit if a user case emerges
  • Obscure filesystems (reiserfs [removed upstream], hfs*, ubifs) — n

Revisit the trim list when a user reports hardware that needs something we dropped.

Per-flavor differences that DON'T go here

These come up in .config discussions but belong in flavor -tuning packages, not the kernel config:

  • Default I/O scheduler per device class (BFQ for rotational, mq-deadline for SSD) → udev rules
  • VM tuning (swappiness, dirty_ratio) → sysctl fragment
  • Network buffer sizes → sysctl fragment
  • CPU governor → tuned profile
  • Forwarding on/off → sysctl fragment
  • vm.max_map_count bumps for gaming → sysctl fragment

See the per-flavor docs (SMOOTHNAS.md, SMOOTHROUTER.md, SMOOTHHTPC.md, SMOOTHDESKTOP.md) for the actual values.

Per-flavor differences that legitimately need kernel support

Currently none. Every user-relevant difference between flavors is solvable in userspace with the canonical kernel. If that changes, this table fills in:

Flavor What they need Why userspace can't do it
— — —

Adding a row is a strong signal to reconsider whether the one-kernel model is still correct.

Updating the config

When a kernel bump introduces new config symbols (common on major bumps):

  1. make kernel-config-update-all refreshes the patched tree, runs olddefconfig, reapplies the SmoothKernel profile, and writes the resulting configs back into configs/.
  2. Review the resulting diff. Sometimes a new option defaults to y and bloats the image; sometimes to n and disables something we want. Check the diff manually.
  3. Run make check-rocm-config before review sign-off; it guards the amd64 amdgpu/HSA symbols that SmoothNAS ROCm plugins need for /dev/kfd. make check-router-config guards, for every arch, the netfilter (inet tables, NAT, flowtable) and PPPoE symbols SmoothRouter needs. The build also runs it against the resolved .config, since olddefconfig drops a symbol whose parents are unset even when the profile asks for it.
  4. Commit the updated configs/smooth-<arch>.config and configs/<kernel-version>/smooth-<arch>.config files as part of the kernel-bump PR.

Out-of-tree modules

Smooth* out-of-tree kernel modules (currently just smoothfs) use the compat.h shim pattern documented in the existing templates. See templates/compat.h.in, DKMS.md, and the update steps in bumping-kernel.md for how to move KERNEL_FLOOR_MAJOR/MINOR across a bump.

New out-of-tree modules in Smooth* repos follow the same pattern: copy the template, wire it into the repo's debian/ packaging, and build via DKMS against the installed SmoothKernel headers.

Non-goals

  • Per-flavor kernel variants. Not shipping linux-smoothkernel-server, linux-smoothkernel-desktop, etc. One binary. See ARCHITECTURE.md.
  • Microarch variants. Not shipping linux-smoothkernel-v3. Possible future addition if demand justifies; not v1.
  • A "configurator" UI for the kernel. Users don't configure kernels. That's what this doc — and one canonical config — exists to avoid.