Skip to main content

ALMA PACKAGE SUPPLY

alma10 and alma10-kitten build separate packages for Albacore and Yellowfin. Intel uses the pinned linux/amd64/v2 child; ARM uses its native linux/arm64 child. Factory RPMs use x86_64 or aarch64. Alma vendor dependencies may use x86_64_v2; that label alone proves no CPU compatibility.

The contract for each target selects its image and publication prefix. Generic packages publish under rpm/<target>/<arch>. For GNOME and XFCE, native families use <family>/<target>-<arch>. Paths for CentOS Stream stay independent.

Evidence and bootstrap​

For each recipe, the build keeps an inventory of packages in its buildroot. It also keeps repositories, dependency transactions, compiler versions, effective flags, and source digests. The policy for compilers preserves the native security flags. It applies the v2 or ARMv8 baseline after the recipe exports its environment. CPU overrides that conflict with this policy fail the build.

For native chains, Mock uses repositories with signatures and an empty repository for local candidates. Each run creates a temporary key to sign candidates. Keep that key outside artifacts. The helper checks each RPM signature before it adds the RPM. It stages metadata with its signature and checks both before replacement.

Hold the repository lock for these changes. Keep the public key and candidate scope with observations of the buildroot. The chain removes private keys on exit.

These keys establish integrity for local candidates. They give no authority to sign production supply. NVRs from prior publication, restored RPMs, and roots from cache cannot substitute for an Alma rebuild. Keep published sources disabled until an authenticated immutable snapshot gives the same evidence.

For CPU inspection, bind evidence to exact RPM bytes and ELF requirements for machine and ISA. Bind each linked library to its provider. If static requirements are absent, execute the exact artifact on a restricted CPU. For vendor inputs, authenticate signatures and evidence for the repository and baseline. Do not invent compiler flags for those inputs.

Inspection always leaves readiness false. Before promotion, the publication gate must verify the producer's authority and a clean transaction for the consumer.

The CPU collector's optional --cache-dir stores only primary metadata in its raw compressed form and vendor RPM blobs, addressed by SHA256. It never caches keys, installed roots or proof verdicts.

The collector fetches the key and repomd.xml afresh on every run. It checks each cached file's type, size and digest against that run's metadata, and rehashes hits and copies. It safely refetches missing, partial or corrupt inputs. If bytes from the network have the wrong digest, collection stops.

Every hit still needs signature checks. The installed ELF files must match their declared byte hashes. Identity checks and tests with restricted-CPU QEMU remain mandatory.

CI restores inputs only from the exact namespace for that cache version. PRs restore their own scope and the trusted base, and save only PR-local inputs. Only trusted default-branch runs save shared inputs. A missing cache slot does not trigger a fallback to production supply.

Run 41144847612 took 1775 seconds for the signed CPU closure across two targets. Its job timestamps do not identify the expensive substeps.

The script now logs phase times within bounded output. The collection.json file includes aggregate times in phaseSeconds for acquisition, index construction, signatures, byte binding and QEMU. These measurements, not the existence of a cache, establish any savings. These results do not measure a speedup for the full collector.

Complete factory selection​

The planner splits selections into deterministic batches if they exceed the three existing matrix shards. Keep slots for continuation of native chains. Include each selected cell once in the original inventory. Do not drop packages to fit a matrix.

The parent dispatches the other batches with the exact source commit and selection digest. Keep the identity of each child run. Each batch and try needs a successful result with the same identity. Coverage stays blocked if a child is absent, incomplete or failed. Batch completion does not establish readiness for package publication.

Diagnosed failures​

SYMPTOM: A genuine v2 build fails on -mtls-dialect=gnu2. CAUSE: The guard for compilers treated every positive -m option as an ISA extension. That included the TLS ABI selector from Alma's native macros. FIX: Allow the exact GNU2 selector. Reject unknown options for TLS and CPU extensions. Tests for the compiler policy include the measured native flags.

SYMPTOM: The parser for candidate signatures accepts a bad signature beside a successful payload digest. CAUSE: Separate checks of the text matched Signature and a different line's OK. FIX: The signature must report success on the same line. Tests for the candidate repository preserve this corruption regression.

SYMPTOM: A symlink in the repository redirects writes during bootstrap. CAUSE: The path resolver hid the symlink before it checked the components. FIX: Reject symlinks before resolution. Keep private state for candidates outside the repository. Keep tests for atomic replacement.