[aledbf]

Where a boot's time goes, measured on the bench fleet, and what that bought: the first SSH login 60 ms sooner and the kernel 6.2 ms faster. Every number below is the bench fleet (AWS c8i.4xlarge, KVM nested), boots interleaved, p50.

What changes in the release

Change	Measured	Run
kernel/patches/0005: a device's PCI capability list is walked once, kept in struct pci_dev	kernel start to "Freeing unused kernel image" 95.1 → 88.9 ms; acpi_init 14 → 10; virtio_pci_driver_init 6.5 → 4.7; inet_init (reads no config space) 7.50 → 7.49; ECAM accesses per boot 1,032 → 680	lab 43327612594 (20 boots)
spin lingers, logind back in the boot	first SSH login over vsock 570/626 → 510/525 ms p50/p95; sshd accepts no later (324 → 321)	lab 43308825692 (10 boots)
No MOTD, pam_motd off in sshd's and login's stacks	a login −7 ms, the first −12 ms	lab 43308825692
sshd@.service waits for the host keys	SSH over vsock and ssh .host worked not at all: the per-connection sshd found no host keys and dropped the connection	boot:ssh-first

Together, the first login from launch should be about 505 ms against 570. That is an estimate, image and kernel measured apart, because the lab boots published releases; the release's own boot:ssh-first and boot:initcalls confirm it.

The console prompt comes about 80 ms later with linger, because spin's manager now starts in the boot. Nothing a workspace does waits for it, but it is what boot:bench's usable column reads.

No fingerprint moves: nothing under machine/ changes, and a resumed guest runs the kernel it was saved with.

New probes

• boot:waits: what waited on what between PID 1 and the login, from systemd's timestamps after the boot: a critical path that accounts for every millisecond, the gaps with no unit activating, and what still starts behind the prompt.
• boot:jobs: PID 1's job queue recorded live over /run/systemd/private by boot/jobwatch, a diagnostic binary in the boot's overlay, beside boots without it. It also records the jobs the first SSH login sets off.
• boot:ssh-first: from the host, over vsock (boot/vsockproxy): when sshd accepts, when the first login is done, and what a second one costs.
• boot:exits: KVM's counters per phase (exits by kind, EPT faults, how the guest's RAM is mapped), and a row without virtio-rng.
• boot:mmio: whose exits they are, from QEMU's trace: every PCI config access by device, every MMIO/PIO access by region.

All five run in the lab (lab.yml).

Fixed tools

• SPIN_GRAPH and SPIN_CPU_SHARE traced nothing since the kernel forced lockdown at confidentiality (1400f6a), which closes tracefs, and reported "the trace is empty". They now print the lockdown level and say how to trace a kernel B built without it (kernel/README.md).
• parseGraph read a function that migrated CPUs (acpi_init opened on CPU 1, closed on CPU 0) as nothing.
• boot:trace takes SPIN_TRACE_APPEND and FLAGS.

Measured and not shipped

• The getty without its wait for dev-ttyS0.device: 342 against 345 ms (lab 43303190733).
• Host keys generated in the boot: 518 against 514 (lab 43312419007).
• virtio-rng removed: −13.6 ms to the prompt (lab 43321809471). That is a fingerprint change, so it is not here.
• pci=noearly: the guest gets no ECAM and does not boot.
• PID 1's ~25 ms before the generators is the BPF verifier checking systemd's socket-bind program (bpf_check). Measured, nothing done.

Checked

• task lint, task test.
• compat.yml on the branch (43329268857): a checkpoint of the release before resumes here; boot:roms; boot:lockdown.
• kernel.yml (43330642231) and image.yml (43330637817) on the branch. image/build.sh asserts on the finished filesystem: no pam_motd, spin lingering, the logind want not overridden, and the sshd@ drop-in.
• report's matrix boots on the patched kernel, every row (lab 43328179961).
• Hotplug on the patched kernel, by hand: spin-machine attach/detach of a disk works, and every device's MSI-X vectors are identical to the release kernel's.
• boot:logind with logind in the boot.

Notes for review

• The commit "Remove motd" also carries linger and logind in the boot (configure-system.sh, optimize-systemd.sh, 10-seats.conf, build.sh, boot/logind_test.go).
• boot:logind's SPIN_LOGIND_NO_SEATS=1 control was written to fail, and on this image (systemd 259) it passes. So the 10-seats.conf drop-in may no longer be needed. Not settled here.
• go.mod gains github.com/godbus/dbus/v5 v5.1.0, used only by boot/jobwatch, which is never in a release.

🤖 Generated with Claude Code