The kernel freezes every process before it loads a hibernated image, so the
screen looked hung for as long as the load took. A drop-in on
systemd-hibernate-resume.service now opens the kernel console just before
resume, draws a band across the middle of tty12 and confines the kernel's
progress lines to it. The running bootvine stops drawing and stops taking the
console back meanwhile, and repaints if no image was restored.
Codeberg rewrites a merged pull request's head ref to refs/pull/N/head once
its branch is deleted, so release-plz never recognised the merged release
pull request and v0.1.0 was never published.
Each sample machine now has a VM check, including new ones for
text-console and scripted-initrd. While a check runs, a recorder takes a
frame every half second through a second QEMU monitor socket, and
record-screencasts encodes the frames into recordings/<machine>.webp with
img2webp, dropping the BIOS frames at other resolutions. The guide shows
each machine's recording in the section it illustrates, replacing the PNG
screenshots. The checks are now named vm-<machine>.
Failed jobs keep their kind: units that broke or timed out get a failure
panel entry with the reason systemd reports (exit status, signal, timeout,
start limit, status text) and their last journal lines read through
libsystemd. Units that failed only through a dependency are dimmed and left
out of the panel, and units that crash after starting count as failed.
Restructure the repository around this:
- move the library and binary into bootvine/ and the scripted demos into
their own unpublished scripted-demos crate, with their VHS tapes
- move the crane builds and NixOS module into nix/
- group each sample VM with its check under vm-tests/, adding text-console,
scripted-initrd, encrypted-volume and failures machines and a vm-failures
check
- add docs/guide.md walking through the screen with every recording, keep
one recording in the README, and re-render all recordings
Since the boot only counted as complete when the default target was
reached, bootvine kept taking its terminal back from a display manager
that started while another unit still held graphical.target. The
StartupFinished signal was never parsed either, so nothing else ended
the watch.
Frames tick every 16 ms; password checks and the bus's fallback wake-ups keep
their own 100 ms clock. Running units show hundredths, and a running bar's tip
fills in eighths when the console font has the block glyphs. The scripted boot
demo uses about 2% of a core.
Console upkeep still runs every second, but a full repaint follows only when it
had to re-quiet the kernel, take the terminal back or noticed a new font, after
reconnecting, and every five seconds as a safety net.
The screen keeps the projected tree, milestones and tally and rebuilds them
after boot changes, or every 100 ms so finished units still fade on time.
Release, 200 units, under load: a frame from the cached projection costs
30-40% of a full frame (0.35-0.53 ms against 1.2-1.9 ms).
Calls send their message and await the reply on the registered socket while
the other streams keep drawing; signals that arrive meanwhile are queued for
the next receive. Restoring systemd's status output moves to a guard that
reconnects on exit, so it no longer depends on where the feed was suspended.
Status modes are an enum instead of raw strings.
Every source is a stream of ScreenInput values: the systemd bus (resolved into
BootChange values), keys, frame ticks, full repaints, password requests and
SIGTERM. They are merged with stream::select and folded by one owned
WatchScreen, drawing once per ready batch. The live boot and the scripted demo
share the fold and differ only in their console desk. tokio current-thread
replaces poll(2) and signal-hook. Release binary: 1.74 MB -> 1.86 MB.
Searching the graph per sibling pair cost O(n^2) searches under a target with
many dependencies. One reachable set per sibling makes the check a lookup.
Release, 200 units: model projection 2.64 ms -> 0.75 ms, whole frame
3.02 ms -> 1.09 ms.