Files
egress-proxies-operator/docs/plans/2026-09-04-1404-kvm-provider.md
T
kacerrandClaude Fable 5.1 bdd1c6ce6c Add plan: KVM/libvirt provider
Design for a kvm provider type: SSH-forwarded libvirt socket, multi-host
config with placement.zone selecting the host, bridged networking with
ARP/guest-agent IP discovery, operator-fetched distro cloud images cached
in the host's storage pool, and a NoCloud seed ISO carrying the CRD's
cloud-init user-data.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
2026-09-10 19:13:51 +00:00

314 lines
17 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Plan: KVM/libvirt provider
**Created:** 2026-09-04 14:04
## Context
The operator provisions egress proxies through named provider instances (today
`gcp` and `kubernetes`). The user has several KVM hypervisors on the home
network and wants proxy VMs created there, using stock Debian/Ubuntu cloud
images so the VM stays a normal machine they can customise through their own
cloud-init user-data. Longer term the operator may become a generic "run a VM
or container with a set of scripts" tool; the KVM provider should not
foreclose that.
The provider contract (`internal/provider/provider.go`: `Create`/`Get`/`Delete`/
`ListByTag`, keyed by `NameFromUID`, cloud-init passed verbatim, GC-by-tag) fits
libvirt with **no CRD or reconciler changes**. All work is a new provider
package plus config/wiring/docs, following the GCP provider's shape.
Decisions taken with the user (2026-09-04):
- **Transport:** SSH to each hypervisor, forwarding the libvirt unix socket
(`ssh.Client.Dial("unix", "/var/run/libvirt/libvirt-sock")` → `libvirt.New(conn)`).
Pure Go, no cgo, no libvirt libs in the image. Verified: `x/crypto/ssh`
`Client.Dial` has a `"unix"` case (`dialStreamLocal`), go-libvirt `New(net.Conn)`.
- **Topology:** one provider instance can span several hosts from day one
(config has a `hosts` list; providerID encodes the host), so adding hosts
later is a config edit with no providerID migration. With one host configured
`placement.zone` may be omitted.
- **Networking:** VM NIC on a host bridge to the LAN, IP from the router's
DHCP. IP discovery via `DomainInterfaceAddresses` (ARP source first, then
guest agent). Provider-injected `vendor-data` installs `qemu-guest-agent`
so agent discovery works on stock images without touching user-data.
- **Base image:** a stock Debian/Ubuntu cloud image (qcow2, cloud-init
preinstalled) that **the operator fetches itself** on first use per host:
`placement.image` names an entry in the provider's `images:` map (url +
optional sha256) or is an `https://` URL directly. The provider streams the
download through libvirt's `StorageVolUpload` into the host's pool, so
nothing is pre-placed and the operator needs no filesystem access on the
host. Per-VM disk is a qcow2 overlay with the cached base as backing store.
- **Seed:** NoCloud ISO (`user-data`, `meta-data`, `vendor-data`) built in
pure Go with `github.com/kdomanski/iso9660` (`WriteTo(w, "cidata")`) and
streamed to the pool with `StorageVolUpload(vol, io.Reader, ...)`. Nothing
to install on the hypervisors.
## Design
### Config (`internal/provider/config.go`)
```yaml
- name: kvm-home
type: kvm
kvm:
sshKeyFile: /etc/proxy-operator/kvm/id_ed25519 # required
knownHostsFile: /etc/proxy-operator/kvm/known_hosts # required unless insecureIgnoreHostKey
insecureIgnoreHostKey: false
diskSizeGb: 10 # overlay capacity; cloud-init growpart expands root fs
vendorData: | # optional override; default installs qemu-guest-agent
#cloud-config
...
images: # optional; placement.image names a key (or is an https URL)
debian-13:
url: https://cloud.debian.org/images/cloud/trixie/latest/debian-13-generic-amd64.qcow2
sha256: "…" # optional; verified after download, re-fetch when it changes
format: qcow2 # default qcow2
ubuntu-24.04:
url: https://cloud-images.ubuntu.com/noble/current/noble-server-cloudimg-amd64.img
machineTypes: # required, ≥1; placement.machineType names a key
small: {vcpus: 1, memoryMiB: 1024}
medium: {vcpus: 2, memoryMiB: 2048}
hosts: # required, ≥1, unique names
- name: nas1 # placement.zone value
address: nas1.home.lan:22 # required
user: virt # required; must be in the libvirt group
socket: /var/run/libvirt/libvirt-sock # default
pool: default # default "default"
bridge: br0 # required
```
Add `KVMConfig`, `KVMHost`, `KVMMachineType`, `KVMImage` structs; a `"kvm"`
case in `validate()` (cross-block checks like the others, required fields
above, unique host names, image URLs must be `http(s)`, sha256 if set must be
64 hex chars). Tests in `config_test.go` extend the existing table.
### Placement mapping
| Placement field | KVM meaning |
|---|---|
| `zone` | host name from `hosts`; empty → the single configured host, else `ErrPermanent` |
| `machineType` | key in `machineTypes`; unknown → `ErrPermanent` |
| `image` | key in `images:`, or an `https://` URL; anything else → `ErrPermanent`. Fetched into the host's pool on first use (see *Image cache*) |
| `region` | ignored |
### providerID
`hosts/<host>/domains/<name>` with `formatProviderID`/`parseProviderID`
(same self-contained style as GCP's `zones/<zone>/instances/<name>`).
### GC contract without labels
Domain XML gets a `<metadata>` child in a private namespace, e.g.
`xmlns="https://gitea.home.hrajfrisbee.cz/kacerr/egress-proxies-operator/kvm"`,
carrying `<managed>true</managed><uid>…</uid><createdAt>RFC3339</createdAt>`.
`Get` reads it via `DomainGetMetadata(dom, DomainMetadataElement, uri)`;
`ListByTag` does `ConnectListAllDomains` on every host and keeps domains where
that call succeeds and `managed == "true"`. `CreatedAt` comes from the metadata
(libvirt has no creation time). A host that is unreachable makes `ListByTag`
return an error (GC skips the sweep that round rather than acting on a partial view).
### Connection management (`internal/provider/kvm/conn.go`)
Per-host lazily dialed `*ssh.Client` + `*libvirt.Libvirt`, cached under a
mutex, re-dialed when `Disconnected()` is closed or a call fails with a
transport error. SSH auth: private key from `sshKeyFile`; host key check via
`knownhosts.New(knownHostsFile)` unless `insecureIgnoreHostKey`. Dial timeout
from `ctx`. The provider calls libvirt through a narrow interface:
```go
type libvirtAPI interface {
DomainLookupByName(name string) (libvirt.Domain, error)
DomainDefineXML(xml string) (libvirt.Domain, error)
DomainCreate(d libvirt.Domain) error
DomainDestroy(d libvirt.Domain) error
DomainUndefineFlags(d libvirt.Domain, f libvirt.DomainUndefineFlagsValues) error
DomainGetState(d libvirt.Domain, flags uint32) (int32, int32, error)
DomainGetMetadata(d libvirt.Domain, typ int32, uri libvirt.OptString, f libvirt.DomainModificationImpact) (string, error)
DomainInterfaceAddresses(d libvirt.Domain, src uint32, flags uint32) ([]libvirt.DomainInterface, error)
ConnectListAllDomains(need int32, f libvirt.ConnectListAllDomainsFlags) ([]libvirt.Domain, uint32, error)
StoragePoolLookupByName(name string) (libvirt.StoragePool, error)
StorageVolLookupByName(p libvirt.StoragePool, name string) (libvirt.StorageVol, error)
StorageVolGetPath(v libvirt.StorageVol) (string, error)
StorageVolCreateXML(p libvirt.StoragePool, xml string, f libvirt.StorageVolCreateFlags) (libvirt.StorageVol, error)
StorageVolUpload(v libvirt.StorageVol, r io.Reader, off, n uint64, f libvirt.StorageVolUploadFlags) error
StorageVolDownload(v libvirt.StorageVol, w io.Writer, off, n uint64, f libvirt.StorageVolDownloadFlags) error
StorageVolDelete(v libvirt.StorageVol, f libvirt.StorageVolDeleteFlags) error
StoragePoolRefresh(p libvirt.StoragePool, flags uint32) error
}
```
(All signatures verified against go-libvirt `main` on 2026-09-04.) A
`hostDialer func(ctx, KVMHost) (libvirtAPI, io.Closer, error)` is the test seam;
`newWithDialer(cfg, dialer)` mirrors GCP's `newWithAPI`.
### Image cache (`internal/provider/kvm/image.go`)
`ensureImage(ctx, host, api, img) (basePath string, err error)` makes the base
image present in the host's pool, once per host per image:
- Volume names: `base-<key>.<format>` for an `images:` key, or
`base-<sha256(url)[:16]>.<format>` for a URL given directly in
`placement.image`. Marker volume `<base>.sha256` (raw, 64 bytes) holds the
digest of the verified upload.
- **Ready check:** base volume exists *and* marker exists *and* (no `sha256`
configured, or marker content == configured `sha256`). Read the marker with
`StorageVolDownload`. Ready → return `StorageVolGetPath(base)`.
- **Not ready:** delete both volumes if present (partial upload from a crash,
or a changed checksum), then `GET url` with the ctx, require a
`Content-Length` (`ErrPermanent` otherwise), `StorageVolCreateXML` the base
with `<capacity unit="bytes">Content-Length</capacity>` and the configured
format, `StorageVolUpload(base, io.TeeReader(body, sha256), 0, 0, 0)` (length
0 = until EOF, so nothing is buffered in the operator), then compare the
digest with the configured one (mismatch → delete base, `ErrPermanent`),
`StoragePoolRefresh` so libvirt re-reads the real qcow2 capacity, create the
marker volume and upload the digest last.
- HTTP 4xx → `ErrPermanent`; 5xx, network, or upload transport errors →
`ErrTransient` (partial volumes are cleaned on the next attempt).
- A per-`(host, image)` `sync.Mutex` map serialises concurrent Creates for the
same image on the same host; the second caller finds it ready.
- **Judgment call:** the fetch runs synchronously inside `Create`, so the first
Proxy for a given image on a given host holds a reconcile worker for the
download (LAN speed, a few hundred MB, once). This keeps `Create` idempotent
and needs no background state. Async fetch with a `Provisioning`-style
progress status is a possible later improvement, noted in the docs.
### Create (idempotent, fire-and-forget after the image is cached)
1. Resolve host, machine type, pool, and image source (`images:` key or URL;
else `ErrPermanent`). `ensureImage` → base path.
2. Overlay volume `<name>.qcow2`: `StorageVolCreateXML` with
`<capacity unit="G">diskSizeGb</capacity>`, `<target><format type="qcow2"/></target>`,
`<backingStore><path>base</path><format type="qcow2"/></backingStore>`.
`ErrStorageVolExist` → reuse.
3. Seed volume `<name>-seed.iso`: build ISO in memory (`user-data` = req.CloudInit,
`meta-data` = `instance-id: <name>\nlocal-hostname: <name>\n`, `vendor-data` =
cfg or default), `StorageVolCreateXML` with exact capacity + `<format type="raw"/>`,
then `StorageVolUpload`. Exists → skip upload (content is deterministic per spec hash;
spec changes recreate the whole instance anyway).
4. `DomainLookupByName(name)`; if `ErrNoDomain` → `DomainDefineXML(buildDomainXML(...))`.
5. `DomainGetState`; if not running → `DomainCreate`. Return providerID.
`buildDomainXML` is a pure function over `encoding/xml` structs: `type="kvm"`,
q35 machine, `host-passthrough` CPU, vcpus/memory from the profile, virtio disk
(overlay) + sata cdrom (seed), `interface type="bridge"` with virtio model,
`org.qemu.guest_agent.0` virtio channel, serial console, virtio rng, no
graphics, and the `<metadata>` block. Unit-tested field by field like
`buildInsertRequest`.
### Get
`parseProviderID` → host → `DomainLookupByName` (`ErrNoDomain` → `ErrNotFound`)
→ `DomainGetState` + `DomainGetMetadata` + `DomainInterfaceAddresses`
(ARP, then agent; first IPv4 wins; agent errors while the agent isn't up yet
are swallowed). Mapping:
| libvirt state | IP | Instance.State |
|---|---|---|
| running | yes | Running |
| running | no | Provisioning |
| shutoff / crashed / paused / pmsuspended | – | Stopped |
| no domain | – | `ErrNotFound` |
### Delete (idempotent)
Destroy if running (`ErrOperationInvalid` on a stopped domain is fine) →
`DomainUndefineFlags(managed-save | snapshots-metadata | nvram)` → delete
`<name>.qcow2` and `<name>-seed.iso` (`ErrNoStorageVol` fine). Volumes are
deleted by name even when the domain is already gone, so a crash between
volume creation and define still gets cleaned up on the next reconcile.
Known gap (documented, not solved now): volumes orphaned with no domain *and*
no CR are invisible to `ListByTag`.
### Errors (`internal/provider/kvm/errors.go`)
`classify(err)` on `libvirt.Error.Code`: `ErrNoDomain`, `ErrNoStorageVol` →
`ErrNotFound`; `ErrNoStoragePool`, `ErrXMLError`, `ErrConfigUnsupported`,
`ErrNoSupport`, `ErrAuthFailed`, `ErrInvalidArg` → `ErrPermanent`; SSH dial /
host-key / transport errors and everything else → `ErrTransient`. Image
download: HTTP 4xx, missing Content-Length, checksum mismatch → `ErrPermanent`;
5xx and network errors → `ErrTransient`. Quota is not detectable from libvirt;
left as transient (revisit with a capacity check later).
### Wiring & manifests
- `cmd/main.go`: add `"kvm": kvm.New` to the builtins map.
- `config/manager/manager.yaml`: mount Secret `kvm-ssh` (optional) at
`/etc/proxy-operator/kvm` (keys `id_ed25519`, `known_hosts`); document the
`kubectl create secret generic kvm-ssh --from-file=...` command.
- `config/samples/providers-config.yaml`: add the `kvm-home` block with the
`images:` map (Debian 13 and Ubuntu 24.04 official cloud-image URLs);
`config/samples/proxy_kvm.yaml`: Proxy with `placement: {zone: nas1,
machineType: small, image: debian-13}` and the same squid cloud-init as the
GCP sample plus `ssh_authorized_keys`.
- `hack/providers-dev.yaml` stays kubernetes-only (needs no secrets).
- Deps: `github.com/digitalocean/go-libvirt` (latest pseudo-version),
`github.com/kdomanski/iso9660 v0.4.0`, promote `golang.org/x/crypto` to direct.
### Docs
- New `docs/kvm-provider.md`: hypervisor prerequisites (user in `libvirt`
group, a bridge to the LAN, a storage pool with enough space), the config
reference including `images:`, how the image cache works (`base-*` and
`*.sha256` volumes in the pool, when a re-download happens, how to evict
with `virsh vol-delete`), the placement table, how to log into a VM
(`ssh_authorized_keys` in user-data, `virsh console` on the host), and the
known gaps (no quota detection, orphan volumes without a domain,
synchronous first-time image fetch).
- Update the provider tables in `README.md`, `docs/architecture.md` §5,
`docs/testing.md` (integration test section).
### Towards the generic runner (not built now, just kept open)
Provider knows nothing about squid. Future hooks: `vendor-data` for injecting an
agent, `QEMUDomainAgentCommand` for reading exit status/output, and a future
`Placement`-independent "job" mode in the CRD. None of this changes the KVM
provider's shape.
## Steps
0. **Save plan + branch.** Copy this file to `docs/plans/<ts>-kvm-provider.md`,
commit alone, then `git checkout -b feat/kvm-provider` off `feat/proxy-operator`
(the working main for this project).
1. **Config.** `KVMConfig` & friends, `validate()` case, table tests.
2. **Pure builders.** `internal/provider/kvm/{domainxml.go,volume.go,seed.go,id.go,metadata.go}`
+ tests: domain XML fields, volume XML, seed contents (unpack the ISO with
the same library's reader and assert the three files + label `cidata`),
providerID round-trip, metadata marshal/parse.
3. **Transport + errors.** `conn.go` (SSH dial, host-key policy, per-host cache,
reconnect), `errors.go` + classification table test. SSH dialer tested
against an in-process `x/crypto/ssh` server that serves a unix socket
forward (cheap, keeps the dial code covered).
3b. **Image cache.** `image.go` + `image_test.go`: `httptest.Server` serving a
small fake image over the `fakeAPI` (which records uploaded bytes);
cases: fresh fetch, already ready, marker missing (partial upload → redo),
checksum changed → redo, checksum mismatch → Permanent + cleanup, 404 →
Permanent, 503 → Transient, two concurrent callers → one download.
4. **Provider methods.** `kvm.go`: `New`, `newWithDialer`, `Create/Get/Delete/ListByTag`
over `libvirtAPI`; `kvm_test.go` with a hand-written `fakeAPI` (domains and
volumes in maps, per-call error injection) covering idempotent re-Create
after each partial step, state/IP mapping, delete-when-gone, ListByTag
across two hosts and with one host down.
5. **Wiring, manifests, samples, docs, CHANGELOG.**
6. **Integration test** (`//go:build integration`, `kvm_integration_test.go`):
skipped unless `KVM_TEST_HOST`, `KVM_TEST_USER`, `KVM_TEST_SSH_KEY`,
`KVM_TEST_BRIDGE`, `KVM_TEST_IMAGE_URL` are set; fetches the image (or finds
it cached), creates, waits for Running+IP, deletes, asserts per-VM volumes
gone and the base image still cached. Deviation from the testcontainers convention,
called out in `docs/testing.md`: libvirt+KVM can't run inside a container
without `/dev/kvm`, so this targets a real hypervisor.
7. **Manual validation** on one real hypervisor, results in
`docs/plans-executions/` (like `docs/gcp-vm-validation.md`), then MR via `tea`.
## Verification
- `make test` green (`-race`, envtest suite untouched but must still pass).
- `go test -short ./internal/provider/...` covers builders, fake-API provider, errors, config.
- Integration: `KVM_TEST_HOST=nas1.home.lan … go test -tags=integration -run TestKVM ./internal/provider/kvm/`.
- End to end: deploy with the `kvm-ssh` Secret and `kvm-home` config against a
host with no cached image, apply `config/samples/proxy_kvm.yaml`, watch the
operator log the image fetch and `virsh vol-list` show `base-debian-13.qcow2`
plus its `.sha256` marker, watch `kubectl get px -w` reach `Ready` with a LAN
IP, apply a second Proxy and confirm no second download, confirm the VM shows up in `GET /v1/proxies` and works with
`hack/` demo egress-IP check, delete the Proxy and confirm `virsh list --all`
and `virsh vol-list default` are clean.