Solaris 9 SPARC: seed-based CI, and raw device enumeration

[danifunker]

Two related pieces of Solaris 9 SPARC work. The release pipeline leg had never run end to
end, and rb-cli could not see a local disk at all.

Part 1 — the release pipeline builds from a seed, not a container

CI no longer builds a toolchain or uses Docker. SOL9_SPARC64_SYSROOT_URL holds a seed —
the cross toolchain, its sysroot, mrustc and both prebuilt standard libraries, 47 MB — which
the job unpacks at / before running scripts/build-sol9.sh.

	Before	After
CI artifact	7 GB image	47 MB seed
CI time	~100 min	~20 min
External deps at build time	Debian base, apt, ftp.gnu.org ×2	one download
Dockerfiles on the release path	3	0

The container rebuilt, every run, an environment that already exists on a developer machine.
It was also a liability: Debian 11 reached EOL mid-development and broke the build twice in
one evening — an expired Release file, then its security packages 404ing out of the pool,
with archive.debian.org carrying no bullseye-security to fall back on.

Two scripts replace the three Dockerfiles: scripts/build-sol9-toolchain.sh builds binutils
and GCC 4.9.4 natively (GitHub's ubuntu-latest and a current desktop are both Ubuntu 24.04
/ GCC 13.3, which 4.9.4 builds under), and scripts/pack-sol9-seed.sh packs the seed.

Verified in a bare ubuntu:24.04 container — eight apt packages plus the seed — and then
in the real pipeline, whose artifact I downloaded and ran on a Sun Blade 2500: it produced a
zstd backup that restored byte-identically with fsck clean.

Part 2 — rb-cli can see local disks

Solaris fell into the catch-all arms of src/os/mod.rs, so show devices listed nothing.
There is now a src/os/solaris.rs beside linux.rs and macos.rs — it has to live there,
because rb-cli-sol9 carries no Rust source of its own and compiles ../src/lib.rs.

Enumeration deliberately opens no device. A wedged driver blocks in open(2) and ignores
O_NONBLOCK — measured on the Blade, where a USB stick behind scsa2usb left processes that
kill -9 could not touch. Solaris' own tools have the same flaw: format(1M) and
rmformat(1) both hang on that disk, and Oracle warns never to point format at a USB
drive. The first version of this module did exactly that, and hung.

So it reads what the drivers publish — the same source iostat -E uses, and the only one
that keeps working: readlink into /devices, /etc/path_to_inst for the driver instance,
then libkstat for Size, Vendor and Product. A device that has stopped answering
reports zero and is skipped. libdiskmgt, which format and the installer use, is 32-bit
only on Solaris 9 and unavailable to a sparcv9 binary.

Only kstat_named_t is modelled; kstat_t stays opaque because kstat_data_lookup returns
a pointer into the library's own buffer, which removes most of the FFI risk.

The whole-disk node differs by architecture and both are probed — SPARC has only SMI slices
and calls slice 2 the whole disk, x86 wraps them in an fdisk table and calls it p0. That is
commented and unit-tested in both spellings, because Oracle's USB docs show p0 throughout
(those examples are x86) and someone would otherwise "fix" the SPARC path to match.

Verified on hardware with the wedged stick still attached — both healthy disks listed
with correct sizes and inquiry strings, the bad one skipped, in 112 ms:

"name": "c0t0d0",  "media_name": "QUANTUM BlueSCSI UltraW",   "bus_protocol": "sd"
"name": "c0t6d0",  "media_name": "BlueSCSI CD-ROM CDU-55S",   "bus_protocol": "sd"

Six latent bugs this flushed out

Each would have fired the first time the pipeline ran for real:

1. SOL9_LIBGCC unset in the container — defaulted to a path that does not exist there, so dist would have refused to package.
2. mksysroot.sh disk mode had never worked — rb-cli get -r lays a directory source out under the destination, producing usr/include/include/stdio.h.
3. That mode left every symlink as a text file, so -lc would find a 12-byte libc.so. Now replayed — 2,400 links on a real disk image, all seven link-time libraries resolving.
4. MRUSTC_TARGET_VER unset in the container path, so mrustc fell back to 1.29 mode and died parsing 1.74's asm!.
5. The cross toolchain was never on PATH for sol9libs, so mrustc used the host x86 gcc and failed with bad value 'v9' for '-mtune='.
6. libc was declared for linux and macos only. Solaris built anyway because rb-cli-sol9 declares it separately, but cargo check --target sparcv9-sun-solaris could not type-check the module — that check is now 24 seconds instead of a 20-minute mrustc round trip.

Known limitations

Tracked in docs/solaris9-open-items.md, which separates known-wrong from untested. The
headlines:

• USB mass storage is beta. scsa2usb wedges on devices that do not implement the full SCSI command set. scripts/solaris-usb-unblock.sh diagnoses it without touching a disk and prints Oracle's remedy, applying it only with --apply.
• The write path still bails for Solaris. Enumeration and read are proven; writing needs a device that is not the root disk.
• A CD-ROM reports removable: no. A failed DKIOCREMOVABLE is treated as "not removable" — known-wrong rather than verified.
• Solaris 10 rests on an untested assumption — forward binary compatibility should make the Solaris 9 artifact run there, but nobody has tried it.

Also simplifies a boolean in src/fs/sfs.rs that a newer clippy flags as nonminimal_bool,
which was failing the pre-commit hook for every commit in the repo.

🤖 Generated with Claude Code