persistMode=backing seeds the persistent OS disk as a thin qcow2 overlay
(`qemu-img create -b <storeImage>`) instead of a full cp, so many VMs share one
base image and each holds only its deltas. Because the overlay reads through a
store path that nix does not otherwise pin (the disk lives outside the store),
and because even a cp'd Windows disk backs onto the store chain, a per-VM
oneshot `vm.vmix-gcroot@<name>` reads the overlay's ACTUAL backing_file at boot
(not the config's current image, which drifts after a rebuild) and symlinks it
under /nix/var/nix/gcroots. It runs before the VM service and outside its
ProtectSystem sandbox. Fires whenever the OS disk is persistent, covering copy
too. Default stays copy, so existing VMs are unaffected.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0117qMyjpuXsjpVAcpJbFD8g
Two defects found while getting a GPU through Proxmox to a nested guest.
Passthrough gave every address its own pcie-root-port, which splits a GPU from
its own HDMI audio: 05:00.0 and 05:00.1 arrived in the guest as two devices on
two buses instead of functions 0 and 1 of one device. Navi needs both halves on
one device to reset or power-manage either, so the guest got a card stuck in D3
and a reset that could not be performed. Addresses are now grouped by
everything left of the function digit, and each group goes behind one root port
at one slot with multifunction=on on function 0 -- which is also where the
VBIOS and the VGA route belong.
The data-disk setup used Initialize-Disk/New-Partition/Format-Volume. Only the
first of those works that early in specialize; the rest need services that are
not up yet, and with ErrorActionPreference=Stop the script gave up straight
after writing a GPT header. The result was a 50G disk carrying 24KB of nothing
and a ProfilesDirectory pointing at a volume that never existed. diskpart works
at that stage. It also tries assigning the letter before laying the disk out,
so a disk that already holds a profile is lettered rather than cleaned.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0117qMyjpuXsjpVAcpJbFD8g
Three things, all in service of putting Proxmox in a VM that can still hand a
GPU to its own guests, and of a Windows VM whose profile survives its OS disk.
pci.viommu.enable emits `-device intel-iommu,intremap=on,caching-mode=on` and
forces kernel-irqchip=split, which interrupt remapping requires. Without an
IOMMU of its own a guest cannot bind a passed-through device to vfio-pci, so
it can never forward one on. The device leads the command line because QEMU
realizes devices in order and intel-iommu must precede what it translates.
pci.vgaPassthrough (default true, so nothing changes for existing VMs) makes
x-vga=on optional. It was forced on the first passthrough device, which is
wrong for a card the guest only forwards onward: it claims the VGA path the
emulated console adapter needs.
customizeImage gains extraDisk, a blank disk attached for the Audit Mode boot
and emitted as the derivation's `data` output. generalize uses it for dataDisk
and profilesDirectory, so the disk is partitioned and the profile relocated
under OOBE in the build VM. That is what removes the need for delayOobeRun --
previously the volume ProfilesDirectory names could not exist until the image
reached real hardware. A second output rather than a directory keeps ${image}
meaning the OS qcow2 for every existing consumer.
generalize also picks up staticIP and profilesDirectory.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0117qMyjpuXsjpVAcpJbFD8g
All wan.net.vmix@* instances created their veth pair with the same
temporary peer name 'vhost' in the host namespace before moving it into
their netns. Parallel starts at boot could steal each other's peer ends,
pairing a host-side vn-<ns> with another namespace's vhost (mismatched
/30s, dead links) or leaving the pair stranded in the host namespace.
Use a per-namespace temporary name (vh-<ns>) and rename to vhost only
after the move, making concurrent creation collision-free.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
NixOS firewall sets conf.all.forwarding=false via mkDefault, which
overrides ip_forward=1. Use normal priority to beat mkDefault.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The NixOS module was importing lib directly with the host's pkgs,
causing image customization to use the host's guestfs-tools instead
of vmix's locked version. guestfs-tools 1.52.2 (from host nixpkgs)
has a bug that overwrites /boot/grub/grub.cfg with resolv.conf
content, breaking VM boot.
Now vmixLib is built once in flake.nix with vmix's own nixpkgs and
passed through the overlay to pkgs.vmixLib. Removes overlay.nix and
module.nix as the logic is inlined in flake.nix.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add nicModel option (default: virtio-net-pci) to allow e1000 for
images without VirtIO drivers
- Restore MAS activation with slmgr /ipk to switch back from IoT
Enterprise S to Enterprise LTSC (which has native RDP server)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When spice.vgamem is set (e.g. 64), uses -device qxl-vga,vgamem_mb=N
instead of -vga qxl (which defaults to 16MB). When null (default),
uses -vga qxl for backwards compatibility.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Only set sandbox = "relaxed" when vmix.namespaces is non-empty.
Safe to import as a default module on all hosts.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- macvtaps working
- only 1 dnsmasq service per namespace
- vms binds to networking services
- lans with domains
- vms no longer assigned same ip (machine id issues)
-
