G1: the current stripe

[tamnd]

This tracks G1, the current stripe. This is the milestone rucc cares about most: it says which of today's kernels GCC 16 builds and boots on the tier 1 platforms, and it proves that rucc-kernel can read gcc-kernel cells as its reference. The full text is in docs/spec/12-milestones.md.

Harness

• gk-boot, the QEMU rig with serial framing and boot budgets
• gk-init for 2.6 and later, built per platform by the era GCC
• Config fragments for eras M13 and M14
• Boots to L6, KUnit to L7, and splat detection for L8
• The three run rule for cells that reach L5, with flaky cells marked
• The i386 platform and its toolchain bundles
• The current stripe of 09.6, dense, on x86_64, i386 and arm64, with defconfig and tinyconfig
• The GCC 16 column, dense over releases 5.0 to 7.x
• The warning census of 11.5 on the Current set
• The config differential across columns of 11.3 on the Current set
• gk publish with the markdown heat maps

Open questions

• 3: the measured cost of a modern cell, and whether tier 2 is cut back
• 8: KVM against TCG on the era diagonal
• 10: the measured cost of allmodconfig on the Current set

Elsewhere

• rucc-kernel runs rk config-diff with a gcc-kernel GCC 16 cell directory as one side, unmodified

Exit criterion

• The current stripe is complete, and every cell has a verdict or a reason for n/a
• rk config-diff in rucc-kernel reads a gcc-kernel cell directory and produces a table, which proves the data contract of 10.4
• reports/new-gcc-16.2.md exists, written as if 16.2 were new, as a rehearsal of 10.8

[tamnd]

#28 adds gk-init for 2.6 and later, plus gk init, which builds it and writes the initramfs.

It's a static program with no libc, using raw syscalls on x86_64, i386 and arm64. The smoke suite has 18 checks, each framed with GK- marker lines. The archive is written by gk with fixed owners, times and inodes, and server2 and server3 produce identical digests for all three platforms.

One deviation from the checklist wording: it isn't built by the era GCC. Each platform pins a GCC column and a host in platforms.toml, which is currently GCC 8.5 in the trixie host, with i386 using the x86_64 bundle and -m32. The 2.6 era columns aren't forged yet, and since every column of a row boots the same archive, a single pinned build is enough. Spec 07.4 now says this. I'll move the pin once older bundles exist.

Next is the QEMU container and gk-boot.

[tamnd]

#29 adds gk boot and the gk-boot QEMU container, so the rig now exists. It boots an image with the platform's initramfs under TCG with 2 CPUs and 1 GB, and writes boot.log and boot.json. It stops on power off, when the budget runs out, or after 5 quiet seconds following the end marker.

As a first test I booted server2's Ubuntu 6.8 kernel, which passed 18 of 18 smoke checks in 15 seconds. That run also caught an init bug that put the wrong suite name on the end line, and it's fixed in the same PR.

I'm leaving the rig item unticked until cells actually boot through it. That is the next PR: L5 and L6 in gk cell, plus the qemu and initramfs coordinates. A 7.2.8 x86_64 image is building on server2 to test it.

[tamnd]

#30 is merged, and with #29 the QEMU rig is done, so that box is ticked.

Cells now boot. gk cell 7.2.8 gcc-16.2.0 --platform x86_64 built in 31 minutes on server2 and reached L6 with 18 of 18 smoke checks. The gk-boot container digest and the initramfs digest are part of a booted cell's identity, and a cell run with --no-boot keeps its old identity.

The three run rule also went in with #30: a cell that boots at all boots three times, takes the lowest rung, and is marked flaky when the runs disagree. I'll tick that box once the arm64 cell running on server2 has gone through it end to end.

A finding from the first 7.2.8 boot: the M11 to M14 fragments build every KUnit suite in, and they all run before init. Under TCG the full set took 328 seconds on 7.2.8 x86_64, more than the 300 second boot budget. Smoke boots now turn KUnit off on the command line, and KUnit gets its own boot and budget in the next PR.

Something to plan for: L7 grades KUnit against the era GCC's cell, and for 7.2.8 that is gcc-14.2.0, which we have not forged. gccs.toml only has the newest point release of each series. Until 12.2.0 and 14.2.0 are forged, cells will record that they had nothing to grade against and stay at L6.

[tamnd]

#31 is merged: KUnit boots and L7 grading.

A cell whose config has KUnit now boots a second time after each smoke boot that passed. That boot runs every suite, with its own budget of 1200 seconds, and kunit.json grades the suites against the era GCC's cell: only suites the era cell passed in all three runs count, and a run reaches L7 when it passed all of them.

On the 7.2.8 x86_64 GCC 16.2 image, the KUnit boot read 176 suites in 283 seconds, and 6 failed, all of them the generic page table iommu suites. Whether that is GCC 16 or the kernel will be clear once the era GCC cell has run. gcc-14.2.0 and gcc-12.2.0 are forging on server2 now for exactly that.

The "Boots to L6, KUnit to L7, and splat detection for L8" box stays open until L8 is in. One thing L8 has to deal with: that KUnit boot logged 19 splats from tests that warn on purpose, so splats on a KUnit boot have to be compared against the reference rather than counted outright.

[tamnd]

#32 is merged, which puts L8 in place: objtool warnings from the build logs and splats from the console. KUnit splats only count when the era GCC's cell did not show the same one. They are matched by a key without the CPU, task and offsets, because the 7.2.8 KUnit boot logs 19 warnings from suites that warn on purpose.

All the rungs are now there in code. I'll tick the ladder box once a cell has climbed past L6 for real, which needs the gcc-14.2.0 era cell. That GCC is forging on server2 now.

Released v0.1.1 with everything from #28 to #32: https://github.com/tamnd/gcc-kernel/releases/tag/v0.1.1

[tamnd]

#34 is merged: gk publish now writes reports/matrix-<platform>.md heat maps, one letter per cell, with a * on flaky cells and the newest cell winning when a square has been run more than once. That ticks the gk publish box. The per-GCC and per-kernel pages come later with gk explain.

On the dev store it gives one row so far, 7.2.8 with R under GCC 16.2 on x86_64. The real tables come with the current stripe.

[tamnd]

#35 adds gk config-diff. It compares the config of the newest cells for two GCC columns, or every pair of neighbouring columns in the store when none are named, and prints a markdown table per pair. I'm leaving the config differential item open until it has been run over the Current set, which needs more than one column per kernel in the store first.

Still running on server2: the arm64 7.2.8 cell (three boots end to end), and forge4 building gcc-14.2.0, the era GCC that 7.2.8's KUnit and splat grading is measured against.