pxe-server/tests/vm/README.md

pxetest VM harness

Proves a new PXE server can be built from this repo, in two libvirt VMs on a laptop, without root and without touching a real network.

Topology

Object	What
pxetest-imaging	libvirt network, bridge virbr-pxetest. No <forward>, no <ip>: the host has no address on it and libvirt runs no DHCP there. Only VM taps are on the bridge.
default	libvirt NAT, 192.168.122.0/24. Stands in for the GE/corp network.
pxetest-server	2 vCPU, 4 GiB, q35 + OVMF Secure Boot (MS keys), 60G thin qcow2 (vda, boot 1), the USB image as a USB disk (sda, boot 2).
pxetest-client	2 vCPU, 2 GiB, OVMF Secure Boot (MS keys), 1G blank disk, network boot first (UEFI IPv4 PXE, virtio-net, OVMF driver, no option ROM). Proves shim -> iPXE -> menu -> WinPE under Secure Boot; WinPE has no virtio driver, so it stops at "Waiting for network".
pxetest-client-nosb	2 vCPU, 2 GiB, OVMF without Secure Boot, e1000e NIC, 64G thin SATA disk. WinPE has in-box drivers for both, so startnet gets a network and maps the shares.

NIC MACs (fixed):

VM	NIC	MAC	Network	Playbook var
server	GE	52:54:00:9e:10:01	default NAT	corp_nic_mac
server	imaging	52:54:00:9e:10:02	pxetest-imaging	pxe_nic_mac
client	-	52:54:00:9e:10:03	pxetest-imaging	-
client-nosb	-	52:54:00:9e:10:04	pxetest-imaging	-

build-image.sh writes $PXETEST_CACHE/site-vars.yml with these two MACs (pxe_nic_mac, corp_nic_mac) and passes it to build-usb.sh --site-vars, so it lands on CIDATA as playbook/site-vars.yml. The installer brings both NICs up with DHCP; the playbook renames them to pxe0 / corp0 and pins pxe0 to 172.16.9.1.

The disk boots before the USB on purpose: a blank disk falls through to the installer, and after the install the server boots its disk instead of re-running autoinstall. CIDATA stays attached for the first-boot playbook.

Never attach a pxetest VM to a physical NIC, a host bridge with physical ports, or macvtap: the server's dnsmasq serves DHCP/PXE as 172.16.9.1. create.sh, start.sh and status.sh check the isolation and refuse to continue if pxetest-imaging ever gains a <forward>, an <ip>, a host IPv4 address or a non-tap bridge port.

Run it

export PXETEST_CACHE=~/.cache/pxetest               # ISO + image (about 10 GB)
export PXETEST_PAYLOAD_ROOT=/home/camp/projects/pxe  # checkout holding offline-packages/, boot-files/, ...
export UBUNTU_ISO_NAME=ubuntu-24.04.3-live-server-amd64.iso  # optional; default 24.04.5

tests/vm/fetch-iso.sh        # the live-server ISO, SHA256SUMS-verified
scripts/download-packages.sh --iso "$PXETEST_CACHE/$UBUNTU_ISO_NAME" "$PXETEST_PAYLOAD_ROOT/offline-packages"
scripts/check-updates.py --debs "$PXETEST_PAYLOAD_ROOT/offline-packages"
tests/vm/build-image.sh      # build-usb.sh --image, upload as volume pxetest-usb.img
tests/vm/create.sh           # network, volumes, both VMs (prints the MAC pins)
tests/vm/start.sh server     # autoinstall, reboot, first-boot playbook
tests/vm/status.sh --shot /tmp/pxetest-shots   # states, leases, screenshots
tests/vm/start.sh client     # once the server playbook is done (Secure Boot client)
tests/vm/stop.sh client
tests/vm/start.sh client2    # Secure Boot off, e1000e: WinPE maps the shares
tests/vm/stop.sh             # ACPI shutdown all, keep them defined
tests/vm/destroy.sh          # remove every pxetest-* object

build-image.sh skips enrollment/, bios-staging/, drivers-staging/ and boot-tools/ by default; --with-enrollment and --with-boot-tools add them back, --size 20G fixes the image size (default auto). If the host has no sfdisk or mtools, the debs are extracted into $PXETEST_CACHE/tools (no install, no root).

Watch the server with virt-viewer -c qemu:///system pxetest-server (VNC on 127.0.0.1) or screenshots. On the installed system, first-boot output goes to /var/log/first-boot.log. Once the GE NIC takes a DHCP lease on the NAT network, status.sh shows its address and ssh pxe@<address> works from the host.

The offline package set must be built against the SAME ISO the image is built from (download-packages.sh resolves against that ISO's base). For a fresh server disk: tests/vm/destroy.sh, then build-image.sh, create.sh, start.sh server.

Memory: server 4 GiB + one client 2 GiB (start.sh refuses a second client). Keep other VMs (win11) off while both run, and stop them when done.

Blancco needs more client RAM than 2 GiB: blancco-init downloads its ~850 MB rootfs into a tmpfs, and at 2 GiB it stops at 48% with "No space left on device". To test it within the 6 GiB budget, shrink the server and grow the client while both are off, and put them back afterwards:

virsh -c qemu:///system setmaxmem pxetest-server 2048M --config
virsh -c qemu:///system setmem    pxetest-server 2048M --config
virsh -c qemu:///system setmaxmem pxetest-client 4096M --config
virsh -c qemu:///system setmem    pxetest-client 4096M --config

Menu keys from the host: virsh -c qemu:///system send-key <vm> KEY_DOWN (separator lines are skipped: Down x1 = Blancco, x3 = Clonezilla, x4 = Memtest), then KEY_ENTER; startnet answers are KEY_ + KEY_ENTER.