Repository navigation
feat: LTS and OGC kernel specs - #67
HuntedRaven7 wants to merge 6 commits into
Conversation
hanthor
left a comment
There was a problem hiding this comment.
I verified the source locks against real upstream — one checks out exactly, the other I could not reach from here. But both specs have defects that mean neither would produce the kernel its name claims, and kernel-centos cannot get past %prep at all.
Source lock verification
kernel-centos — downloaded the real SRPM (161,067,765 bytes) and hashed it:
ACTUAL: 99599fa0d9105f6207e903e98cf8add7f6d4c63da4e1500951e9ba52fd5ed5e7ca6e844a6f0bc69b248964da8fd8026c555142e93f2f300f3e6eef31f8c38665
CLAIMED: 99599fa0d9105f6207e903e98cf8add7f6d4c63da4e1500951e9ba52fd5ed5e7ca6e844a6f0bc69b248964da8fd8026c555142e93f2f300f3e6eef31f8c38665
Exact match. Good.
One caveat on that URL: mirror.stream.centos.org/10-stream/BaseOS/source/tree/Packages/ is a rolling directory. It currently lists kernel-6.12.0-{254,257,260,264,266}.el10.src.rpm; -264 is no longer the newest and older builds get pruned from that tree over time. This pin will 404 eventually. vault.centos.org or the Koji URL would be stable.
kernel-ogc — tag v7.2.3-ogc3 exists, at 151eb3adafae8820f349a4454c1f1e8d5fc991a4 (git ls-remote). I could not verify its sha512: the archive fetch is blocked from this session (HTTP 403 from the egress proxy, not from GitHub). So that lock is unverified here and someone should confirm it independently.
A structural concern with it regardless: https://github.com/.../archive/refs/tags/*.tar.gz is a GitHub auto-generated archive. GitHub does not guarantee those are byte-stable — their gzip settings have changed before and silently invalidated checksums across whole ecosystems. Every other entry in config/upstream-sources.json that I looked at points at a real upstream release tarball. A release asset, or a locally-produced archive published as a lookaside, would be safer for something as expensive to rebuild as a kernel. Also note the repo has already moved on: v7.2.3-ogc4, v7.2.4-ogc1 and v7.2.4-ogc2 all exist.
kernel-centos cannot complete %prep
Release: 264.el10%{?dist}
...
%prep
rpm2cpio %{SOURCE0} | cpio -idmv *.tar.xz
tar -xJf linux-*.tar.xz
cd linux-%{version}-%{release}
%{release} expands to the whole Release field including %{?dist} — so on Fedora 44 that cd target is linux-6.12.0-264.el10.fc44.
I unpacked the real SRPM header to check what the directory is actually called. The payload contains three archives:
tar.xz names in SRPM header: ['kernel-abi-stablelists-6.12.0-264.el10.tar.xz',
'kernel-kabi-dw-6.12.0-264.el10.tar.xz',
'linux-6.12.0-264.el10.tar.xz']
so the extracted tree is linux-6.12.0-264.el10, with no dist suffix. The cd fails, %prep exits non-zero, the build dies. This has never been built. (%{version}-264.el10 literal, or %{version}-%{?!dist:%{release}} style, or simply cd linux-6.12.0-264.el10.)
Neither spec builds the kernel it is named after
%build
make mrproper
make defconfig
make -j$(nproc)
make defconfig generates the generic upstream x86_64 defconfig. For kernel-centos that discards the entire CentOS Stream configuration and every downstream patch — the SRPM's .config files and its patch series are extracted and then ignored. What comes out is a vanilla 6.12 kernel wearing a CentOS name and NVR. Same for kernel-ogc: the "gaming-focused patches" are in the tree, but the OGC config that makes them meaningful is not applied. A defconfig kernel also lacks a great deal of what a bootc image needs (many filesystems, most storage/network drivers, and the module signing and BTF settings the distro config sets).
make mrproper before make defconfig also deletes anything a real %prep would have staged, which is consistent with the config never being used.
Other issues in both specs
- No
%{?dist}-independent NVR and noProvides: kernel. Nothing can depend on these, anddnfwill not treat them as kernels. - No
kernel-devel/kernel-headers/kernel-modulessubpackages. Anything in the factory that needs kernel headers to build gets nothing. /lib/modules/*/is owned by a bare glob with nodepmodin%postand no%ghosthandling ofmodules.depand friends. Installing this leaves an unusable module tree.- No
BuildRequires: dwarves(pahole), required forCONFIG_DEBUG_INFO_BTF, which anything using BPF needs. License: GPLv2is a retired Fedora identifier; current guidelines want SPDX,GPL-2.0-only.make -j$(nproc)instead of%make_build, so the builder's job settings and flags are bypassed.%changelogentries are* Tue Sep 08 2026 Robin - ...with no email address;rpmbuildwarns and some lint gates reject it.
Factory-level concerns
- Nothing in
config/runtime-contract.tomlorconfig/bluefin-packages.tomlreferences either package, so these are two multi-hour build jobs added to every matrix run for something no image currently consumes. The PR body says "for future Utah feature support" — if that is the intent, a stage assignment and a note about who will consume them would help, or they could wait until there is a consumer. - The PR body says "
just checkpasses with 195 source RPMs", and that reproduces — merged intomainlocally I getvalidated 195 source RPMs,fc=0,72 passed. But that is only the inventory-shape check. It confirms the JSON is well formed and the Packit config was regenerated; it does not compile anything, which is why the%prepbug survived it.
What I would need to change my mind
- Fix the
cdpath inkernel-centos. - Apply the real distro configuration rather than
defconfig— forkernel-centosthat means using the configs and patch series from inside the SRPM, which is most of the work in a real kernel spec. - Independently confirm the
kernel-ogcsha512, and consider a stabler source than a GitHub auto-archive. - Add
Provides: kernel, the-devel/-modulessubpackages anddepmodhandling, or say explicitly that these are staging placeholders not meant to be installed. - Prove at least one of them builds. Nothing here has been through
rpmbuild.
The inventory plumbing (source locks, .hummingbird-upstream.json, .packit.yaml regeneration, updated count assertions in the four test files) is all done correctly — that part is clean. It is the specs themselves that need work.
Generated by Claude Code
- Fix kernel-centos %prep cd path (use literal 264.el10, not %{release})
- Use distro configs from SRPM instead of defconfig
- Add Provides: kernel, kernel-devel, kernel-headers, kernel-modules
- Add kernel subpackages (devel, headers, modules) with proper %files
- Add %post/%postun with depmod for module handling
- Add BuildRequires: dwarves for BTF support
- Change License to SPDX GPL-2.0-only
- Use %make_build/%make_install macros
- Fix changelog email format
- Use vault.centos.org for stable source URL
Assisted-by: Nemotron-3-Ultra via GitHub Copilot
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
|
Updated it I would love another review on this, |
|
Closing as stale: conflicts with Generated by Claude Code |
Adds two new kernel packages for future Utah feature support.
kernel-centos: CentOS Stream 10 LTS kernel (6.12.0-264.el10)kernel-ogc: Open Gaming Collective kernel (7.2.3-ogc3)Both specs follow the existing factory conventions: one spec per package directory, SHA-512 source locks in
config/upstream-sources.json, Packit monorepo config regenerated, and tests updated for the new inventory count.Source locks:
mirror.stream.centos.org/10-stream/BaseOS/source/tree/Packages/kernel-6.12.0-264.el10.src.rpmgithub.com/OpenGamingCollective/linux/archive/refs/tags/v7.2.3-ogc3.tar.gzValidation:
just checkpasses with 195 source RPMs