Repository navigation
Releases
v1.111.1-20261031
VXVisor Fabric, signed cluster traffic v2, migration network fix, installer fixes (Preview)
VXVisor Fabric: multi-host L2/L3 networking (new Fabric page)
- What it is. One logical switch across hosts, built on BGP EVPN and VXLAN. A wizard takes you through
hosts → underlay → access → a live preflight → deploy. You then add port groups that every host in the
fabric carries. Design and status are indocs/FABRIC.md. - L2 and L3. Port groups are either routed or L2-only.
- Routed: an anycast gateway on every host and a VRF per org, so traffic between port groups is
routed on the host where the VM runs. DHCP gives static leases, and every host answers the same way. - L2-only: a plain stretched segment.
- Routed: an anycast gateway on every host and a VRF per org, so traffic between port groups is
- Who can talk to whom. A connectivity table drives a distributed firewall that runs on every host.
Traffic between port groups is denied by default; traffic within one is allowed. - Egress. A port group reaches the outside through a tenant router connected on the Fabric page, or
through an address you enter. - Guards.
- On routed port groups a guest can use only its leased MAC and IP.
- Guests can't run a DHCP server, send IPv6 router adverts or send VLAN-tagged frames.
- Broadcast and multicast floods are rate-limited.
- VXLAN is accepted only from fabric hosts.
- Two underlays.
- A dedicated NIC or VLAN. Fabric MTU 1550, so guests keep 1500.
- Over the management network. Each host's tunnel endpoint rides an encrypted WireGuard tunnel
(UDP 51871) between the hosts' static management addresses.- The management interface is only read: its MTU, addresses and routes are never changed.
- Tunnel traffic is marked low priority (DSCP CS1).
- Guests get MTU 1390.
- It's refused on a DHCP management address or a cluster VIP.
- Replication. Every fabric change needs every member host to agree, or it isn't made. A member that
missed a change catches up on its own. A dead host can be evicted. - Health. The hosts table shows state, BGP and BFD sessions, duplicate-address flags, VXLAN guard drops
and WireGuard handshakes. - Status. Tested with unit tests, a two-namespace lab and live single-host tests. It hasn't been run
across real hosts yet (seedocs/ROADMAP.md).
Live migration keeps the VM's network
- Bug fixed. A live-migrated VM used to come back with an unbridged NIC on the destination, so it was
offline. - Port group by identity. The VM's port group now travels as org + name + VLAN + VNI + subnet. The
destination attaches its own matching port group, and the guest's address is reserved there. - Precheck. The move is refused before anything is copied when the destination has no match or more
than one.
Cluster traffic signing v2
- What's covered. Node-to-node requests are now signed over the method, path, query, body and a
single-use nonce, and every response is verified. Streamed disk copies are checked before they're
committed. - Admin proxy. The admin proxy between nodes now signs the user too.
- Why. Before, a captured request could be replayed with a different body, or a response forged, by
anyone on the cluster network.
Routing
- One FRR config. Host BGP and the fabric's EVPN are composed into one FRR config, which FRR checks
before it replaces the live file. A config FRR rejects never touches the running one. - NAT table swap. The NAT table is swapped in one checked transaction.
- BGP guard. New guard
inet vdxbgp: inbound BGP (TCP 179) is accepted only from configured neighbours
and fabric hosts.
Installer
- Package installs fixed. Since mid-October a fresh install had silently skipped every "where
available" package: FRR, LXC, ZFS/LVM, NFS/CIFS/Ceph/iSCSI, lldpd, nbdkit and wireguard-tools. - RustGuac console built and configured. The installer now builds and sets up the RustGuac console
(browser SSH/RDP/VNC/SPICE) from fixed, known-good source versions, so the console page's RustGuac button
works on a fresh host.- rustguac and its guacd listen on loopback only.
- Browsers reach them through vdxd's session-checked proxy on
<bind-ip>:8769. - The first install takes several minutes longer and needs access to GitHub and crates.io.
- wireguard-tools. Now installed, for the fabric over management.
Upgrade notes
- Every cluster node must run this version. The old signing is refused, so upgrade all nodes together.
Migration between nodes on different versions is refused. - Re-run
deploy/install.shonce on existing hosts. The in-app update replaces only the binary. The
installer brings the packages the old bug skipped, wireguard-tools and the RustGuac console. It leaves an
existing[guac]config section alone. - Inbound BGP. Peers that aren't configured on the Host Routing page are now refused.
- Database. DB migration 60 (
fabric) is additive.
v1.109.1-20261030
Snapshots and instant clones, CBT warm import, Veeam backup API (Preview)
Snapshots for every disk type
- Every backend. Snapshot, revert and delete now work on zvol, LVM thin and Ceph RBD disks, as well as qcow2.
- Consistency. On a running VM with a guest agent, the guest's filesystems are frozen and then thawed, giving an application-consistent snapshot. Without an agent the snapshot is crash-consistent. Each snapshot is labelled with which kind it is.
- UEFI boot settings. The UEFI NVRAM is saved with each snapshot and restored on revert.
- ZFS revert. A ZFS revert to an older snapshot asks for confirmation before it discards newer snapshots. It only proceeds if those newer snapshots are VXVisor's, unlocked, and have no clones.
- Instant clone. You can create a new instance from a snapshot, or from "now".
- The clone gets a new ID, MAC addresses, NVRAM, console ports, address lease and cloud-init identity.
- zvol and RBD clones share blocks with their origin. Purging the origin makes them independent first.
- Deleting a snapshot that clones depend on is refused, and the error names those clones.
- Snapshot policies (new page under Storage).
- A policy covers a VM, a folder or a whole tenant.
- It runs hourly, daily or weekly, and keeps the last N snapshots and/or D days' worth.
- An optional immutable lock means no one can delete those snapshots, or revert over them, before they expire. This applies to instance admins too.
- Tenant admins manage only their own org's policies.
- REST API. Snapshot and clone endpoints are in the REST API.
CBT warm migration from VMware (opt-in)
- The checkbox. The VMware import dialog now has a "Use CBT (warm migration)" checkbox. It appears only when VDDK is installed.
- What ticking it does. Ticking it is your consent for VXVisor to turn on Changed Block Tracking on that one source VM and to take short-lived
vxvisor-cbt-*snapshots.- The disks copy while the source keeps running.
- Only changed blocks sync after that, every 30 minutes by default or when you press Sync now.
- Cut over… can shut the guest down through VMware Tools, runs a final sync, and then creates the instance.
- Snapshot safety. VXVisor only removes snapshots it created and recorded, including after a crash. Your own snapshots are never touched.
- CBT setting. CBT is restored to how it was before.
- Read-only when unticked. Unticked, the import stays strictly read-only on the source, as before.
- Status. Tested against VMware's simulator and with real nbdkit. It has not yet been tested against a real vCenter.
Veeam Backup & Replication via an oVirt-compatible API (experimental, off by default)
- How Veeam sees it. VXVisor can present itself to Veeam's oVirt KVM plug-in as an oVirt engine.
- What it provides.
- The engine REST subset Veeam uses: SSO, inventory, VMs, disks, backups, checkpoints, image transfers, and restore from OVF.
- An imageio-compatible disk transfer service.
- Incremental backups of qcow2 disks, using QEMU dirty bitmaps and pull-mode backup.
- Support for Veeam's worker VMs.
- Turning it on. Enable it with an
[ovirt]config block:enabled=true. It runs on its own TLS listeners (443 and 54322) on the management IP, with a per-install CA. Only instance admins can sign in, and every change is audited. - Status. It passed the oVirt Python SDK and imageio client conformance tests. It has not yet been tested against a real Veeam server. Details are in
docs/VEEAM-OVIRT-API.md.
Other changes
- SPICE password. SPICE consoles now require the per-launch console password, the same one VNC uses, instead of running without one.
- NBD client. There is a new shared NBD client (
internal/nbd), used by the CBT import and the backup API.
Behavior changes
- Moves with snapshots. Moving or migrating an instance that has snapshots is refused until they're deleted. Before, a move silently lost them.
- Live qcow2 snapshots. Live qcow2 snapshots now save the disk only, not memory. The old
savevmfailed on UEFI VMs, and revert always needed the VM stopped. - Revert after a backup. Reverting a qcow2 snapshot drops that disk's backup checkpoints, so its next Veeam backup is a full one.
- Clones of a TPM guest. A clone of a TPM guest gets an empty TPM, so BitLocker will ask for its recovery key.
Upgrade notes
- DB migrations 53–59 are additive.
- The Veeam API stays off unless you enable it in config.
v1.108.2-20261029
Networking audit hardening (Preview)
A full audit of host switching, tenant networking, L3/routing and VM/container attachment, with every finding fixed and re-checked on a live host. Full design notes: docs/TENANT-NETWORKING.md §32.
Security and isolation
- Validation at two layers. Every networking input is checked in the service layer and again where it's written out (dnsmasq, FRR, NAT and firewall rules, LXC config, VXLAN units).
- SLIRP guard. New nft table
inet vdxslirpstops user-NAT guests from reaching host loopback services. - VXLAN guard.
inet vdxvxlanaccepts UDP:4789 only from configured peers. - Tenant L2 guard.
bridge vdxtenantl2drops rogue DHCP server replies, router advertisements and BPDUs from untrusted ports. - Tenant routes. A tenant gateway exports only its own downlinks: there is an export prefix-list, OSPF is passive on downlinks, and neighbors must be on connected networks.
- Tenant VMs and containers. Tenant-org VMs never run without a port group, and containers attach to port groups rather than raw bridges.
Management and robustness
- Management detection fails closed when the listener is a wildcard and there is no default route.
- VLAN filtering on every vSwitch. Each vSwitch now runs
vlan_filtering=1. At boot, vSwitches, vdxd-made teams and set MTUs are rebuilt. - In-use checks before delete. A NIC, team, vSwitch or port group that is still in use can't be claimed or deleted. Instances in Trash count as in use.
- Routing and FRR changes are applied first and saved only on success, with rollback on failure.
- Physical NICs not managed by netplan still come up for carrier, but without IPv6 autoconf.
Workflow and naming (closer to NSX)
- Nav renamed. The Networking nav group is now Switching, Gateways, Firewall, External Connectivity and Host Routing.
- NIC teams can be edited in place (mode, hash, members), and a non-management vSwitch can have its MTU set.
- Tenant networks can be edited and made the org default. VLAN IDs are reserved so tenants don't collide.
- Topology diagram. A NATed bridge such as
lxcbr0now shows its real egress (NAT → <default-route interface> via <members>), and a container port group's link runs level into its vSwitch.
Behavior changes to know
- Frames a guest tags itself are dropped on a vSwitch port: VLANs come from port groups.
- VLAN 1 is refused for a port group.
- DHCP needs a gateway.
- Tenant networks are IPv4 only and can't overlap host subnets.
- Containers need a port group.
- Trunk names can be at most 10 characters.
Upgrade notes
- DB migrations 49–52 are additive.
- Restarting
nftables.serviceflushes the vdxd tables. Restart vdxd afterwards to rebuild them.
v1.106.2-20261028
Preview v1.106.2 — tenant separation, tenant vSwitches with VLAN port groups and routing out to the network, and a host-isolation security fix.
Tenant view
- An instance admin who switches into a tenant org now sees exactly that tenant: no host NICs, bonds, vSwitches, routing, Tenant Egress, storage backends, cluster or settings. The same service-layer checks the tenant's owner gets enforce this; the pages don't just hide it. Only the org quota and resetting a member's password stay available, marked "provider".
- Fixed host metrics leaking to tenants: the Overview host gauges and disk/network strip, and Monitoring's host live and history data.
Tenant vSwitches and port groups (migration 48)
- A tenant network is now the tenant's vSwitch. Tenants see a vSwitches card: add a vSwitch, pick its uplink, connect or disconnect. A newly created uplink is connected to the tenant's default vSwitch automatically.
- + Port group adds VLAN 2–4094 port groups to a vSwitch. VLAN numbers are private to each tenant and never clash across tenants. A port group is either:
- standalone: its own host gateway and DHCP, reachable only by its own VMs.
- routed: no gateway; a tenant router is its gateway.
- carried out the uplink: the provider's network gives it its gateway and DHCP.
- L2 out: an access uplink carries one port group untagged, chosen with Set native. QinQ and dedicated uplinks also carry every other port group with its VLAN tag.
- L3 out: on vRouters, + Attach port group makes a router a port group's gateway, and + Uplink (WAN) gives the router an address on the uplink, a default route to the upstream gateway and optional NAT.
Security
- The host no longer routes for tenant networks. Before this, a tenant VM on a network with a host gateway could be routed by the host onto the management LAN and reach SSH and the admin UI. From a tenant network, the host now answers only DHCP and pings to that network's own gateway address.
Fixes
- netns routers work on Ubuntu 26.04. AppArmor stopped the routing daemons (staticd/bgpd/ospfd) from starting in a router's own folder, so routers got no addresses and no routes.
install.shnow installs a small AppArmor addition that allows the same files one folder down; the profiles stay in enforce mode. Routers also now startmgmtd, which FRR 10 needs for static routes. - Disconnecting an uplink no longer loses its VLAN tag (reconnecting used to build a broken interface).
- Deleting a router with a port-group or uplink leg no longer deletes the tenant's vSwitch.
install.shnow creates/var/lib/vdxclone/tmp, theTMPDIRthe systemd unit sets.
Behavior changes
- When an uplink carries a port group, that port group's host gateway and DHCP are switched off: the host must not hold an IP or hand out DHCP on the provider's network. They come back if the uplink is disconnected or a different native VLAN is chosen.
- Tenant DHCP no longer offers the host as a DNS server; VMs get 1.1.1.1 and 8.8.8.8 through DHCP.
Upgrade notes
- Re-run
deploy/install.shonce on existing Ubuntu 26.04 hosts (or installdeploy/apparmor/frr.d-vxvisoras/etc/apparmor.d/abstractions/frr.d/vxvisorand reload the FRR profiles withapparmor_parser -r). The in-app update replaces only the vdxd binary, and tenant routers won't work without this. - Migration 48 is applied on first start (additive, forward-only). Rolling back with
vdxd.bakis only safe to a build that already includes it. - The desktop client is unchanged; it is rebuilt at v1.106.2 to keep the unified version, so connected clients will offer the update.
v1.105.0-20261028
Preview v1.105.0 — everything since v1.99.0: Tenant Networking v2 + v3, the optional RustGuac console, and a VM-lifecycle fix.
Tenant networking v2 (migration 45)
- Per-org admins self-serve their org's logical network surface, fully isolated: segments (transactional VNI identity allocator), standalone or routed/gatewayless, attachable to a per-org netns vRouter.
/vroutersand/vfirewallare now available to org admins (transit links and LXC routers stay instance-admin). vRouter names are unique per org instead of globally.- Physical layer (NICs, bonds, vSwitch uplinks, host routing, transit links) stays instance-admin only.
Tenant networking v3 (migrations 46–47)
- Tenant Egress console (
/trunks, instance admin): designate external trunks, define per-tenant VLAN pools, and map each tenant's virtual uplink onto the wire asaccess(802.1q),qinq(802.1ad, tenant inner VLANs preserved) ordedicated(whole NIC). - Shared management-uplink trunks: a trunk can ride the management bond, access-only, and never modifies the parent link (no MTU or offload changes).
- "Disable tenant egress" tears everything down and restores trunks byte-for-byte (crash-safe).
RustGuac advanced console (optional, off by default)
- Browser SSH/RDP/VNC/SPICE via a local rustguac, fronted by vdxd auth through a cookie-gated proxy listener. Enabled only by a
[guac]config block. No DB migration. rustguac is a separately installed add-on, not part ofinstall.sh.
Fixes
- Exited VMs no longer linger as zombie processes. vdxd never reaped QEMU children, so a guest that powered itself off (or a crashed QEMU) kept looking alive until vdxd restarted: the VM stayed "running" in the UI, Stop always waited out its full ~13s timeouts, and a migration receiver that died at launch wasn't detected. QEMU is now reaped on exit, and a per-launch log file-descriptor leak is fixed too.
Upgrade notes
- Migrations 45, 46 and 47 are applied on first start (additive, forward-only). Rolling back with
vdxd.bakis only safe to a build that already includes them. - The desktop client is unchanged since v1.99.0. It is rebuilt at v1.105.0 to keep the unified version, so connected clients will offer the update.
v1.99.0-20261025
Preview v1.99.0 — Tenant Networks v1 (per-org ISOLATED networks: dedicated bridge, no uplink, optional gateway+DHCP; default-isolated for new orgs; cross-tenant VLAN collision guard; RESTRICT-FK-protected teardown). Also: netns vRouter ApplyLegs is now re-apply idempotent (latent 2nd-leg bug fixed). Routed/L2/vRouter external modes + segment allocator are the next phase (docs/TENANT-NETWORKING.md).
v1.98.0-20261025
Preview v1.98.0 — dashboard 'I/O pressure' now shows real busiest-disk utilization (io_ticks), not PSI io which over-reported ~85% on a healthy host.
v1.97.0-20261025
Preview v1.97.0 — new Preview baseline after promoting v1.96.0 to Stable (dev->main). Same code as Stable v1.96.0; carries the Preview channel ahead so future changes can be staged/tested.
v1.96.0-20261025
Preview v1.96.0 — Windows self-update helper no longer flashes a console window (CREATE_NO_WINDOW alone; DETACHED_PROCESS removed — the two are mutually exclusive and together caused the visible cmd/ping window).
v1.95.0-20261025
Preview v1.95.0 — bump to exercise the Windows client self-update now that the dialog-result fix (v1.94) is in place.