fix(scripts): make bench-minvmd-boot safe beside other checkouts, and n>1 actually work - #1023

Three bugs in scripts/bench-minvmd-boot.sh, all found while using it to A/B a new guest rootfs. Two made the bench actively destructive or useless; the third made it unrunnable on macOS.

1. Teardown killed other checkouts' VMs

teardown() { pkill -9 -f "__krun-vmm" 2>/dev/null; … }

That pattern matches every checkout's VM. Running the bench beside a parallel working copy SIGKILLs that copy's live dev stack — sessions and all — between each of N runs, and once more on exit.

It matched a sibling checkout's 44-minute-old VM:

$ pgrep -f "__krun-vmm"
30474 /Users/norrie/code/minimal/target/debug/minvmd __krun-vmm   ← different checkout

scripts/reap-vms.sh already documents the rule this violates:

Matching is scoped to THIS checkout's binaries […] a bare-name pkill would only add collateral — an unrelated checkout's live VM, or podman's gvproxy.

So the fix applies that existing discipline rather than inventing one. Both the supervisor and the __krun-vmm grandchild re-exec via current_exe(), so both carry the binary's absolute path in their cmdline. $BIN is resolved to an absolute path first, because it is conventionally passed relative (./target/debug/minvmd) and that would match nothing.

2. Only the first run was a real sample

Every run after the first reported FAILED (no vm-up).

teardown SIGKILLs the VM, which leaves the bound ssh.sock behind in the state dir. libkrun registers that path with listen=true on the next boot, so it hangs in krun_start_enter until the 12 s timeout. With one state dir shared across runs, n=10 really meant n=1.

Each boot now gets a fresh state dir. That is also the right fidelity — this is a cold-boot benchmark, and a reused state dir carries the previous run's data volume, which is not the state a cold boot starts from.

It additionally isolates the bench from any other checkout: minvmd refuses to boot when the default state dir has a VM registered, and that dir is shared across checkouts even though the binaries are not. A sibling dev stack alone was enough to fail every run with minvmd is already running.

3. mktemp -d -t blew the socket path limit

-t honours TMPDIR, which on macOS is a ~50-byte /var/folders/... path. The daemon socket beneath it then exceeds the 103-byte sun_path limit and the boot dies before it starts:

socket path /var/folders/…/providers/local-minvmd0/ssh.sock is 118 bytes, exceeds the 103-byte limit

A short /tmp template avoids it, matching what scripts/session-e2e.sh already does.

Verification

shellcheck clean at 0.11.0 and at CI's 0.10.0.

Exercised end-to-end on the job it was written for — boot-to-READY across two guest rootfs images, 10 runs each, with a sibling checkout's stack up throughout:

All 20 samples valid, distributions non-overlapping, and the sibling checkout was still running afterwards. Before this PR that comparison was not obtainable: run 1 of each set was the only sample, and the bench would have killed the sibling stack 22 times.

🤖 Generated with Claude Code