v0.9.1

Bugfixes

• Domains are now undefined with flags VIR_DOMAIN_UNDEFINE_NVRAM and VIR_DOMAIN_UNDEFINE_TPM. These defaults may change in the future and be part of a domain block like create and delete. (#1203 )
• Fix SIGSEGV when connecting to a hypervisor over qemu+ssh. Fixes #1210 (#1211)
• Misc documentation fixes #1209, #1210

Features

Support for full libvirt API (XML) (#1208 )

The provider now supports the whole libvirt API 🥳 (* that is supported by libvirtxml), thanks to a code generation engine which generates the whole terraform glue for the schemas and conversions.

For now, the usual resources (domain, network, volume, pool) are included, but this opens the door to handle other resources (secrets, etc) with little effort.

Migration Guide: 0.9.0 → v0.9.1

⚠️ as the schema is now generated, the documentation is now injected into the code generation. As there is no machine readable documentation for libvirt XML, we generated a set of documentation metadata using AI. This process can be improved over time.

Due to the introduction of the generator and some bugs in the 0.9.0 schema, we had to do some changes in the schema.

This document explains how to move Terraform configurations from provider v0.9.0 (the last manual schema) to the current HEAD that uses the libvirt-schema code generator. It only covers resources and attributes that existed in 0.9.0: domains, networks, storage pools, and storage volumes. Anything new that HEAD exposes can simply be added following the generated schema documentation.

What Changed Globally

1. Attr names now mirror libvirt XML – the generator emits snake_case names derived from the XML schema (e.g., accessmode → access_mode, portgroup → port_group). Set exactly the fields you care about; anything left null stays absent in the XML.
2. Value/unit pairs are explicit – whenever libvirt exposes a value with a unit attribute the provider now has two attributes (memory + memory_unit, capacity + capacity_unit, etc.). Leaving the unit unset lets libvirt use its default.
3. Presence/"yes"/"no" semantics follow libvirt – booleans that previously toggled simple structs may now expect yes/no strings when libvirt models them as attributes (e.g. os.loader_readonly). True presence booleans (like features.acpi) still use Terraform bools.
4. Nested objects match the XML tree exactly – device sources, interfaces, backing stores, etc. now use the full nested structure. Plan to touch every place where v0.9 flattened things like source.pool or filesystem.source.
5. Metadata is structured – string blobs became { metadata = { xml = <<EOF ... } } so we can extend later without breaking state.

Domain Resource

Top-level attribute mapping

v0.9 attribute	HEAD attribute(s)	Notes
unit	memory_unit	Same semantics, renamed so every value/unit pair is consistent.
max_memory	maximum_memory	Value only; use maximum_memory_unit if you previously used a non-default unit.
max_memory_slots	maximum_memory_slots	Same semantics.
current_memory	current_memory + optional current_memory_unit	Value stays the same; set the unit explicitly if you relied on non-default units.
metadata (string)	metadata = { xml = <<EOF ... EOF }	Wrap your XML in the nested object.
os.arch	os.type_arch	The type_* prefix mirrors <os><type arch="..."/>.
os.machine	os.type_machine	Same rationale as above.
os.kernel_args	os.cmdline	Field name matches the XML <cmdline> element.
os.loader_path	os.loader	0.9 kept the loader path in a separate attribute; now it is the element’s value (see “value + attributes” below).
os.loader_readonly (bool)	os.loader_readonly (string)	Accepts "yes"/"no" because the XML attribute is a string.
os.nvram.*	os.nv_ram = { file, template, format = { type = ... } }	Rename plus richer structure.
devices.filesystems[*].accessmode	access_mode	All camelCase names were converted to snake_case.
devices.filesystems[*].readonly	read_only	Same semantics.
devices.interfaces[*].source.portgroup	source = { network = { port_group = ... } }	See below for the full source mapping.
devices.rngs[*].device	backend = { random = "/dev/urandom" } or backend = { egd = { ... } }	Backends are now modeled exactly like the XML.

OS block specifics

• os.boot_devices is still a list, but if you previously stored strings you now provide objects: boot_devices = [{ dev = "hd" }, { dev = "network" }].
• Loader/read-only/secure/stateless flags now accept the literal XML strings ("yes"/"no"). Wrap them in tostring() if you had boolean locals.
• NVRAM becomes os = { nv_ram = { file = "/var/lib/libvirt/nvram.bin", template = "/usr/share/OVMF/OVMF_VARS.fd", format = { type = "raw" } } }.

Loader value + attributes

<loader> is a “value + attributes” element. The path is the value (os.loader), and every XML attribute becomes a sibling attribute:

os = {
  loader          = "/usr/share/OVMF/OVMF_CODE.fd"
  loader_type     = "pflash"
  loader_readonly = "yes"
  loader_secure   = "no"
  loader_format   = "raw"
}

Leave the attribute unset to let libvirt pick its default (the provider preserves user intent for optional attributes).

Disks and filesystems

0.9 flattened every disk source. HEAD requires you to pick the XML variant explicitly:

# v0.9
source = {
  pool   = libvirt_pool.test.name
  volume = libvirt_volume.test.name
}

# HEAD
source = {
  volume = {
    pool   = libvirt_pool.test.name
    volume = libvirt_volume.test.name
  }
}

# File-based disk (previously source.file)
source = { file = "/var/lib/libvirt/images/disk.qcow2" }

# Block device
# Block device
source = { block = "/dev/sdb" }

Filesystems follow the same pattern. Replace the old flat fields with nested objects:

# v0.9
filesystems = [{
  source     = "/exports/share"
  target     = "shared"
  accessmode = "mapped"
  readonly   = true
}]

# HEAD
filesystems = [{
  source = { mount = { dir = "/exports/share" } }
  target = { dir = "shared" }
  access_mode = "mapped"
  read_only   = true
}]

Variant notation

Every <source> element with mutually exclusive children (files, volumes, blocks, etc.) becomes an object whose attributes map 1:1 to the libvirt XML children. Only set the branch you need:

# Filesystem backed by a block device
source = { block = { dev = "/dev/vdb" } }

# RAM-backed filesystem with extra attributes
source = { ram = { usage = 1024, unit = "MiB" } }

Even if a variant has additional attributes in XML, the generated struct exposes them in that nested object (e.g., ram = { usage = 1024, unit = "MiB" }). This pattern is consistent across disks, filesystems, host devices, etc.

Interfaces

source.network, source.bridge, and source.dev are now mutually exclusive nested objects. Example conversions:

# Network-backed NIC (v0.9)
source = { network = "default" }

# HEAD
source = { network = { network = "default" } }

# Direct/Macvtap NIC (v0.9)
source = { dev = "eth0", mode = "bridge" }

# HEAD
source = { direct = { dev = "eth0", mode = "bridge" } }

portgroup became port_group, wait_for_ip stays the same helper object.

RNG / TPM / other devices

• RNG devices now mirror <backend>. Use backend = { random = "/dev/urandom" } for /dev/random or backend = { egd = { source = { mode = "connect", host = "unix", service = "..." } } } for EGD sockets.
• TPM backends are nested (backend = { emulator = { path = "/var/lib/swtpm/sock" } }). Map your previous backend_type to one of the backend objects: emulator, passthrough, or external.
• Graphics, consoles, serials, and video devices already used nested objects in 0.9; the only change is snake_case attribute names (auto_port, websocket, etc.).

Metadata

0.9 stored raw XML as a string. Now wrap it:

metadata = {
  xml = <<EOFXML
<libosinfo:libosinfo xmlns:libosinfo="http://libosinfo.org/xmlns/libvirt/domain/1.0">
  <libosinfo:os id="http://libosinfo.org/linux/2024" />
</libosinfo:libosinfo>
EOFXML
}

Storage Volume Resource

Key differences:

v0.9 attribute	HEAD attribute(s)	Notes
format (string)	target = { format = { type = "qcow2" } }	The format lives under the target block now.
permissions.*	target.permissions.*	Same keys, just explicitly under target.
backing_store.format	backing_store = { format = { type = "qcow2" } }	Mirrors libvirt <format> element.
capacity	capacity + optional capacity_unit	Leave capacity_unit unset to keep KiB.
allocation	allocation + allocation_unit (read-only)	Useful when libvirt reports GiB/MiB units.
path (computed)	still path, but it mirrors target.path	You no longer set this manually. Use pool target paths to control locations.

Everything else (name, pool, create/content) behaves exactly like 0.9. Plan/apply will touch terraform state automatically once you update the config.

Storage Pool Resource

The generated schema simply fills in additional optional sub-objects (source.host, source.auth, features, etc.). All attributes that existed in 0.9 keep their names and shapes:

• target = { path = "/var/lib/libvirt/pools" } works unchanged.
• target.permissions.* still take strings, not integers.
• source.device = [{ path = "/dev/sdb" }] keeps the same structure.

Unless you opt into the new nested fields you do not need to change existing pool configurations.

N...

v0.9.0

⚠️ ⚠️ ⚠️ ⚠️ This version of the provider breaks compatibility ⚠️ ⚠️ ⚠️ ⚠️

Background

When this provider was developed, the idea was to mimic a cloud experience on top of libvirt. Because of this, the schema was done as flat as possible, features were abstracted and some features like disks from remote sources were added as convenience.

The initial users of the provider were usually makers of infrastructure software who needed complex network setups. Lot of code was contributed which added complexity outside of its initial design.

So for long time I wanted to restart the provider under a new design principles where:

• HCL maps almost 1:1 to libvirt XML and therefore almost any libvirt feature can be supported
• Most of the validation work is left to libvirt which is already doing it
• No abstractions or extra features, and when they do, they should be designed in a way that are quite independent
• More consistency, for example, most libvirt APIs can be mostly separated in lifecycle (create, destroy) which map quite well to terraform resources, and then some query APIs, which map well to data sources. This was not the case how we implemented for example querying the IP addresses
• No unnecessary defensive code, for example, checking that a volume exist when referenced, is a problem that terraform solves if the ID is interpolated and libvirt solves with its own checks, if the volume is referenced by a hardcoded strings.

I knew 1.0 would never come in the current form.

The new provider

The new provider is based on the new plugin framework. This gives us some room for better diagnostics and better plans.

It makes definitions more verbose, but it also means we can implement any libvirt feature. Defaults work as long as they are defaults in libvirt.

Migration plan

You can find the legacy provider in the v0.8 branch. New releases can be done of 0.8.x versions to add bugfixes, so people who rely on it have a path forward. I'd likely not maintain much of 0.8.x, but I guess many people will help here, as they do today with different PRs.

There is no automated way of migrating the HCL of previous providers, but given that it is documented how the new schema is defined, which was not the case with the previous schema, it should be much easier to drive LLMs to perform a conversion.

You should check the documentation and README, which will give you an idea of the main differences and equivalences, but here is an example of the new schema to get an idea:

terraform {
  required_providers {
    libvirt = {
      source = "dmacvicar/libvirt"
    }
  }
}

provider "libvirt" {
  uri = "qemu:///system"
}

# Base Alpine Linux cloud image stored in the default pool.
resource "libvirt_volume" "alpine_base" {
  name   = "alpine-3.22-base.qcow2"
  pool   = "default"
  format = "qcow2"

  create = {
    content = {
      url = "https://dl-cdn.alpinelinux.org/alpine/v3.22/releases/cloud/generic_alpine-3.22.2-x86_64-bios-cloudinit-r0.qcow2"
    }
  }
}

# Writable copy-on-write layer for the VM.
resource "libvirt_volume" "alpine_disk" {
  name     = "alpine-vm.qcow2"
  pool     = "default"
  format   = "qcow2"
  capacity = 2147483648

  backing_store = {
    path   = libvirt_volume.alpine_base.path
    format = "qcow2"
  }
}

# Cloud-init seed ISO.
resource "libvirt_cloudinit_disk" "alpine_seed" {
  name = "alpine-cloudinit"

  user_data = <<-EOF
    #cloud-config
    chpasswd:
      list: |
        root:password
      expire: false

    ssh_pwauth: true

    packages:
      - openssh-server
    timezone: UTC
  EOF

  meta_data = <<-EOF
    instance-id: alpine-001
    local-hostname: alpine-vm
  EOF

  network_config = <<-EOF
    version: 2
    ethernets:
      eth0:
        dhcp4: true
  EOF
}

# Upload the cloud-init ISO into the pool.
resource "libvirt_volume" "alpine_seed_volume" {
  name = "alpine-cloudinit.iso"
  pool = "default"

  create = {
    content = {
      url = libvirt_cloudinit_disk.alpine_seed.path
    }
  }
}

# Virtual machine definition.
resource "libvirt_domain" "alpine" {
  name   = "alpine-vm"
  memory = 1048576
  vcpu   = 1

  os = {
    type    = "hvm"
    arch    = "x86_64"
    machine = "q35"
  }

  features = {
    acpi = true
  }

  devices = {
    disks = [
      {
        source = {
          pool   = libvirt_volume.alpine_disk.pool
          volume = libvirt_volume.alpine_disk.name
        }
        target = {
          dev = "vda"
          bus = "virtio"
        }
      },
      {
        device = "cdrom"
        source = {
          pool   = libvirt_volume.alpine_seed_volume.pool
          volume = libvirt_volume.alpine_seed_volume.name
        }
        target = {
          dev = "sdb"
          bus = "sata"
        }
      }
    ]

    interfaces = [
      {
        type  = "network"
        model = "virtio"  # e1000 is more compatible than virtio for Alpine
        source = {
          network = "default"
        }
        # TODO: wait_for_ip not implemented yet (Phase 2)
        # This will wait during creation until the interface gets an IP
        wait_for_ip = {
          timeout = 300    # seconds, default 300
          source  = "any"  # "lease" (DHCP), "agent" (qemu-guest-agent), or "any" (try both)
        }
      }
    ]

    graphics = {
      vnc = {
        autoport = "yes"
        listen   = "127.0.0.1"
      }
    }
  }

  running = true
}

# Query the domain's interface addresses
# This data source can be used at any time to retrieve current IP addresses
# without blocking operations like Delete
data "libvirt_domain_interface_addresses" "alpine" {
  domain = libvirt_domain.alpine.name
  source = "lease" # optional: "lease" (DHCP), "agent" (qemu-guest-agent), or "any"
}

# Output all interface information
output "vm_interfaces" {
  description = "All network interfaces with their IP addresses"
  value       = data.libvirt_domain_interface_addresses.alpine.interfaces
}

# Output the first IP address found
output "vm_ip" {
  description = "First IP address of the VM"
  value = length(data.libvirt_domain_interface_addresses.alpine.interfaces) > 0 && length(data.libvirt_domain_interface_addresses.alpine.interfaces[0].addrs) > 0 ? data.libvirt_domain_interface_addresses.alpine.interfaces[0].addrs[0].addr : "No IP address found"
}

# Output all IP addresses across all interfaces
output "vm_all_ips" {
  description = "All IP addresses across all interfaces"
  value = flatten([
    for iface in data.libvirt_domain_interface_addresses.alpine.interfaces : [
      for addr in iface.addrs : addr.addr
    ]
  ])
}

Feedback is appreciated. There will be a long journey for people to port and iron all the issues, but it is clear this is the path to go.

Docs: https://registry.terraform.io/providers/dmacvicar/libvirt/latest/docs

v0.8.3

• Fix for ssh bug #1165

Full Changelog: v0.8.2...v0.8.3

v0.8.2

What's Changed

Content sniffing

• The provider no longer detects the image format qcow2 using content sniffing for remote HTTP images. If you leave it blank, it will just set it based on the extension. This allows to use HTTP servers without HTTP Range support.

• Add ability to set mode=private to a macvlan interface. by @farsonic in #1154

Upgrade dependencies

• Bump golang.org/x/crypto from 0.27.0 to 0.31.0 by @dependabot in #1138
• Bump golang.org/x/net from 0.29.0 to 0.33.0 by @dependabot in #1157

Bug fixes

• Bugfix: ssh port override for #1116 by @memetb in #1117
• fix(ci): failing terraform fmt (#1158) by @dmacvicar in #1159
• fix(domain): restore error handling for network operations by @SJFCS in #1144
• fix: fix the wrong error return value by @cangqiaoyuzhuo in #1161

New Contributors

• @SJFCS made their first contribution in #1144
• @farsonic made their first contribution in #1154
• @cangqiaoyuzhuo made their first contribution in #1161

Full Changelog: v0.8.1...v0.8.2

v0.8.1

What's Changed

This release is mostly about fixes for the SSH transport, which was released with many bugs in v0.8.0

• Do not panic on invalid SSH key by @scabala in #1103
• Sshconfig missing bugfix - addresses issue #1105 by @memetb in #1109
• allow for multiple default identity key files by @memetb in #1112

Experimental LVM storage pool support

There is a new experimental feature, support for LVM storage pools. I don't use myself this type of pools, so I put together all the contributions and made the code ready for release mostly based on integration tests. Try it and give feedback.

• feat: lvm storage pool impl by @jimnydev in #1088

New Contributors

• @scabala made their first contribution in #1103
• @jimnydev made their first contribution in #1088

Full Changelog: v0.8.0...v0.8.1

v0.8.0

What's Changed

Two big features include improved ssh config support (for example for supporting jump hosts) and a new data source for host information.

• expanded ssh_config parameters for qemu+ssh uri option by @memetb in #1059
• feat: add data sources to extract node and device information by @muresan in #1042

Breaking changes

• DNS is enabled by default, like in libvirt. #1100
• Wait intervals for polling libvirt are reduced, making everything faster (including testsuite)

Other highlights:

• Acceptance testsuite is finally fully passing again
• Many code cleanups
• Updated golangci-lint
• Many updated dependencies
• Mark disk wwn and nvram arguments as computed by @wfdewith in #1064
• Default machine by @e4t in #1014
• Add Combustion resource to use instead of the ignition one by @cbosdo in #1068

Community

We activated discussions, so that the community can share useful files, help each other and also get announcements.

• https://github.com/dmacvicar/terraform-provider-libvirt/discussions

Contributors

Thanks to all the community for their contributions and for supporting other users:

• @muresan made their first contribution in #1042
• @wfdewith made their first contribution in #1064
• @kubealex made their first contribution in #1056
• @shafer made their first contribution in #927
• @testwill made their first contribution in #1086
• @memetb made their first contribution in #1059
• @michaelbeaumont
• @cbosdo
• and others... (let me know if I missed you)

Full Changelog: v0.7.6...v0.8.0

v0.7.6

Features

• initial ssh config file support (#933 )

Thanks @jbeisser 🥳

v0.7.5

Fixes

• Fix for configuring network when guest agent is not ready (#1037)
• Make IP address configuration more robust by not stopping prematurely (#1048)
• build with go 1.21

Special thanks to @rgl , @pstrzelczak 🙏

v0.7.4

This release was done to fix the expired GPG key (#1035)

v0.7.2

Fixes

• upgrade ingition dependency
• port to the new libvirt-go dialer constructor
• make 'option_value' for dnsmasq optional (#960)
• Fix malformed connection remote name when using ssh remote uri (#1030)
• Fix test make target to run all tests (#1034)
• Update URL to show how to setup cert (#1007)

Thanks to contributors @michaelbeaumont @flat35hd99 @tiaden @e4t