Skip to content

Raze Linux v1.0.0 - #1

Merged
SwashX-dev merged 7 commits into
mainfrom
develop
Nov 5, 2026
Merged

SwashX-dev merged 7 commits into
mainfrom
develop

Conversation

@SwashX-dev

Copy link
Copy Markdown
Owner

First release of Raze Linux.

What's included

  • r6 — Rust init + single-process service supervisor (PID 1 verified in Raze Core; matches s6 activation latency with 2 processes instead of 27)
  • Rix — Rust package-management front-end with native, Nix-compatible NAR hashing, store verification, generations/rollback and a 27× faster search; installs/builds delegated to Nix
  • Distribution — NixOS-derived, pinned nixpkgs, Linux 7.3.1, sudo-rs, niri desktop with automatic software-rendering fallback, kernel profiles incl. a custom minimal kernel (−16 % boot, −30 % memory vs generic in VMs)
  • Installer — Calamares with Raze branding and a flake-generating install module
  • Docs, benchmarks, CI — architecture/installation/performance docs, real screenshots, tests/report.json, GitHub Actions

Validation

  • Unit/integration: r6 (14 + 18), rix (16) — host, Nix sandbox, GitHub Actions
  • VM: checks.core-boot (r6 PID 1), checks.desktop — local and GitHub Actions
  • Full graphical install of the release ISO in QEMU/KVM (UEFI) with boot-from-disk, login, network, package install, sudo-rs, reboot/persistence, poweroff; live and installed sessions without 3D

Not included (documented)

Rift compositor; r6 as default PID 1 of installed systems (systemd; experimental r6 entry not boot-tested); native Nix evaluator; BIOS/manual-partitioning/dual-boot/real-hardware testing.

🤖 Generated with Claude Code

…ging and CLI

r6-init (PID 1): mounts API filesystems, reaps orphans, runs rc.init/rc.shutdown,
respawns the supervisor with backoff and falls back to an emergency shell; drives
ordered shutdown, unmount and reboot(2). Subreaper mode when not PID 1.

r6-supervise: one event loop (poll + signalfd) supervises all services:
dependency-ordered parallel start, s6-style readiness notification, restart
policies with exponential backoff and rate limiting, stop timeouts with SIGKILL
escalation, dependents-first stop, reload, persistent enable/disable, stale
instance cleanup after a supervisor crash, size-rotated logs.

r6-activate validates service graphs (cycles, unknown deps) and prints the
parallel start plan. r6 is the control CLI.

Tests: 13 unit + 18 end-to-end tests run the real binaries unprivileged.
Rix (native Rust): nix32 encoding, NAR serialisation and hashing (verified
bit-for-bit against nix hash path), store path validation, profile
generations with atomic rollback, profile manifest and flake.lock readers,
a cached package index for instant search, and store verification.
Delegated to Nix (and printed as such): evaluation, builds, substitution,
profile add/remove/upgrade, flake updates, garbage collection, nixos-rebuild.

Distribution:
- Pinned nixpkgs (nixos-unstable 1d060086) with Linux 7.3.1
- sudo-rs replaces sudo (execWheelOnly, use_pty, short timestamp)
- Kernel profiles: generic, performance (preempt=full + scx_bpfland),
  legacy (6.18 LTS), minimal (custom defconfig + kvm_guest build)
- Desktop selection (niri default; rift, hyprland, dwl, xfce, lxqt, plasma,
  gnome, dwm) with greetd/tuigreet for Wayland tiling sessions
- r6 service definitions validated at build time; optional r6 PID 1 boot entry
- Raze-branded Calamares with a flake-generating install module
- Raze Core: busybox + r6-init initramfs image with an s6 comparison tree
- VM checks: Raze Core boot and Raze Desktop integration
…e Core boots with r6 as PID 1

- r6-init: kill(-1, 0) also succeeds for kernel threads, so the final
  shutdown wait always ran to its 3 s timeout. Scan /proc for userspace
  processes instead (PF_KTHREAD). Measured poweroff: ~3.2 s -> ~0.2 s.
- Stale-instance cleanup also sends SIGCONT, like a normal stop.
- The supervisor-crash test asserts on PIDs instead of a shell trap
  (busybox ash may die from a group SIGTERM before running its trap) and
  never leaks r6-init on failure.
- Raze Core: single root tree, decompressed virtio + af_packet modules,
  HUP for the console shell, consistent PID 1 timestamps for r6 and s6.
- checks.core-boot passes: r6 PID 1, service restart, DHCP, logs, poweroff.
…s s6 benchmark

- r6-init waits for rc.init via the signalfd instead of 10 ms polling.
- Per-service progress messages require r6.verbose (or R6_VERBOSE):
  synchronous serial-console writes were the largest r6 boot cost.
  PID 1 -> services started: 44.6 ms -> 32.3 ms (median, 10 boots).
- benchmarks/core-init.py boots Raze Core with r6-init and with an
  equivalent static s6-svscan PID 1 (same kernel, initramfs, services),
  interleaved, and reports latency, RSS, PSS, process counts.
- Static s6/execline build fix (missing pkg-config) for a like-for-like test.

Results (median of 10): pid1->services r6 32.3 ms, s6 32.4 ms;
processes 2 vs 27; RSS 1778 vs 3172 KiB; PSS 1766 vs 1223 KiB.
…er ISO; desktop VM test passes

Installer
- Lock /etc/nixos before nixos-install. nixos-install hashed the flake,
  then wrote flake.lock into it -> 'NAR hash mismatch'. Found by a real
  Calamares run in QEMU; the remaining install path was validated manually
  (install, reboot from disk, login, rix install, sudo-rs, reboot, persistence).
- Rift is no longer offered until it meets the release criteria.

Desktop
- raze-session: niri refuses software EGL renderers (black screen in VMs
  without 3D). raze-gpu-check probes the GBM renderer; without hardware
  acceleration a themed sway session with the pixman renderer is started.
- foot [colors] -> [colors-dark]; ASCII fuzzel prompt; absolute wallpaper
  path; visible live-ISO installer button. niri, foot, fuzzel and sway
  configs are validated at build time.

Size
- Disable speech-dispatcher by default (~700 MiB of voices), drop the
  NixOS default font set incl. CJK (~150 MiB), gvfs without samba (~90 MiB).

Tests
- checks.desktop: sudo-rs auth, branding, dbus-broker, zram, NetworkManager,
  greetd autologin into the software fallback session, waybar/swaybg,
  rix hash/verify against nix. Passes.
- tests/qemu: QMP, VNC and guest-agent helpers for driving installer VMs.
- raze-gpu-check is now a 16 KiB C program (GBM + EGL surfaceless GLES
  context, GL_RENDERER) replacing mesa-demos' eglinfo (61 MiB).
- wiremix (Rust PipeWire mixer) replaces pavucontrol, whose GTK4 stack
  pulled in gstreamer-bad and flite (~60 MiB).
- ISO squashfs uses xz: zstd -15 gave 2.40 GB, over the 2 GiB GitHub
  release-asset limit.
- Calamares is wrapped instead of rebuilt; rix honours RIX_NIXPKGS.
- Raze Minimal kernel fragment: add BPF_SYSCALL for CGROUP_BPF, drop
  PREEMPT_VOLUNTARY (model chosen at boot with PREEMPT_DYNAMIC); fix quoting.
- docs/: architecture (overview, kernel, r6, rix, rift), installation
  (graphical, manual, recovery), development, performance methodology,
  known issues. CONTRIBUTING, CODE_OF_CONDUCT, SECURITY, CHANGELOG.
- GitHub Actions: rust-ci, qemu-tests, kernel-build, benchmarks,
  iso-build, release (ISO size gate, checksums).
ISO
- 2.11 GB (1.961 GiB), under the 2 GiB release-asset limit: xz + x86 BCJ,
  no separate reference closure (its unique part was an initrd that installs
  rebuild anyway), locale list instead of the all-locales archive, no gparted.
- Validated end to end in QEMU/KVM (UEFI): live boot with and without 3D,
  full Calamares install, boot from disk, login, network, rix install,
  sudo-rs, services, reboot, persistence, poweroff. sha256
  51412a1d886aa91349a94306d85746476d21b2a651a391d8c906da99b0698188

Kernel
- Raze Minimal v2: RAID6 benchmark off, ATA modular, no netconsole, no
  DEBUG_WX. Kernel start -> PID 1 360.9 -> 169.6 ms (generic 201.4 ms);
  memory 34.9 MB vs 50.2 MB generic.

Docs and data
- README, benchmarks (r6 vs s6, kernel profiles, rix search 27x vs nix
  search, ISO size history), real screenshots, tests/report.json (17 pass,
  0 fail, 2 not run), release notes.
@SwashX-dev
SwashX-dev merged commit 9f3ccb3 into main Nov 5, 2026
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant