Files
podman/pkg/machine/hyperv/stubber.go
lstocchi cfa1b51775 fix hyperv ignition cleanup and error handling
this fixes two bugs in Hyper-V VM ignition handling:

1. Changed `err := readAndSplitIgnition()` to
`err = readAndSplitIgnition()` so the deferred cleanup function can
properly detect errors and clean up ignition
2. The function was trying to remove keys with pattern `vm.ElementName + index`, but
the actual keys were created using `"ignition.config." + index` pattern.

It also improve the deletion by calculating in how many parts the ignition file is splitted when being added in the registry and delete just them instead of blidnly iterating through 0-50.

Signed-off-by: lstocchi <lstocchi@redhat.com>
2026-01-21 15:50:17 +01:00

748 lines
21 KiB
Go

//go:build windows
package hyperv
import (
"bytes"
"errors"
"fmt"
"os"
"os/exec"
"path/filepath"
"github.com/Microsoft/go-winio"
gvproxy "github.com/containers/gvisor-tap-vsock/pkg/types"
"github.com/containers/libhvee/pkg/hypervctl"
"github.com/containers/libhvee/pkg/kvp/ginsu"
"github.com/containers/podman/v6/pkg/machine"
"github.com/containers/podman/v6/pkg/machine/define"
"github.com/containers/podman/v6/pkg/machine/env"
"github.com/containers/podman/v6/pkg/machine/hyperv/vsock"
"github.com/containers/podman/v6/pkg/machine/ignition"
"github.com/containers/podman/v6/pkg/machine/vmconfigs"
"github.com/containers/podman/v6/pkg/machine/windows"
"github.com/containers/podman/v6/pkg/systemd/parser"
"github.com/sirupsen/logrus"
"go.podman.io/common/pkg/strongunits"
)
type HyperVStubber struct {
vmconfigs.HyperVConfig
}
func (h HyperVStubber) UserModeNetworkEnabled(_ *vmconfigs.MachineConfig) bool {
return true
}
func (h HyperVStubber) UseProviderNetworkSetup() bool {
return false
}
func (h HyperVStubber) RequireExclusiveActive() bool {
return true
}
func (h HyperVStubber) CreateVM(_ define.CreateVMOpts, mc *vmconfigs.MachineConfig, builder *ignition.IgnitionBuilder) error {
var err error
callbackFuncs := machine.CleanUp()
defer callbackFuncs.CleanIfErr(&err)
go callbackFuncs.CleanOnSignal()
hwConfig := hypervctl.HardwareConfig{
CPUs: uint16(mc.Resources.CPUs),
DiskPath: mc.ImagePath.GetPath(),
DiskSize: uint64(mc.Resources.DiskSize),
Memory: uint64(mc.Resources.Memory),
}
// Allow creation in these two cases:
// 1. if the user is Admin
// 2. if the user has Hyper-V admin rights and there is at least one *NEW* machine.
// *NEW* machines are those created with the vsock entry having the ToolName field.
//
// This is to prevent a non-admin user from creating the first machine
// which would require adding vsock entries into the Windows Registry.
if err := h.canCreate(); err != nil {
return err
}
// count number of existing machines, used later to determine if Registry should be cleaned over a failure
machines, err := h.countMachinesWithToolname()
if err != nil {
return err
}
// Callback to remove any created vsock entries in the Windows Registry if the creation fails
removeRegistryEntriesCallBack := func() error {
// Allow removal only if user is Admin and this is the first machine created.
// If there are already existing machines, the vsock entries should remain.
//
// There is no need to check for admin rights here as this is already a requirement
// to create the first machine and so it would have failed earlier.
if machines > 0 {
return nil
}
if err := vsock.RemoveAllHVSockRegistryEntries(); err != nil {
return fmt.Errorf("unable to remove hvsock registry entries: %q", err)
}
return nil
}
callbackFuncs.Add(removeRegistryEntriesCallBack)
// Attempt to load an existing HVSock registry entry for networking.
// If no existing entry is found, create a new one.
// Creating a new entry requires administrative rights.
networkHVSock, err := vsock.LoadHVSockRegistryEntryByPurpose(vsock.Network)
if err != nil {
if !windows.HasAdminRights() {
return ErrHypervRegistryInitRequiresElevation
}
networkHVSock, err = vsock.NewHVSockRegistryEntry(vsock.Network)
if err != nil {
return err
}
}
mc.HyperVHypervisor.NetworkVSock = *networkHVSock
// Add vsock port numbers to mounts
err = createShares(mc)
if err != nil {
return err
}
netUnitFile, err := createNetworkUnit(mc.HyperVHypervisor.NetworkVSock.Port)
if err != nil {
return err
}
builder.WithUnit(ignition.Unit{
Contents: ignition.StrToPtr(netUnitFile),
Enabled: ignition.BoolToPtr(true),
Name: "vsock-network.service",
})
builder.WithFile(ignition.File{
Node: ignition.Node{
Path: "/etc/NetworkManager/system-connections/vsock0.nmconnection",
},
FileEmbedded1: ignition.FileEmbedded1{
Append: nil,
Contents: ignition.Resource{
Source: ignition.EncodeDataURLPtr(hyperVVsockNMConnection),
},
Mode: ignition.IntToPtr(0o600),
},
})
vmm := hypervctl.NewVirtualMachineManager()
err = vmm.NewVirtualMachine(mc.Name, &hwConfig)
if err != nil {
return err
}
vmRemoveCallback := func() error {
vm, err := vmm.GetMachine(mc.Name)
if err != nil {
return err
}
return vm.Remove("")
}
callbackFuncs.Add(vmRemoveCallback)
err = resizeDisk(mc.Resources.DiskSize, mc.ImagePath)
return err
}
func (h HyperVStubber) Exists(name string) (bool, error) {
vmm := hypervctl.NewVirtualMachineManager()
exists, _, err := vmm.GetMachineExists(name)
return exists, err
}
func (h HyperVStubber) MountType() vmconfigs.VolumeMountType {
return vmconfigs.NineP
}
func (h HyperVStubber) MountVolumesToVM(_ *vmconfigs.MachineConfig, _ bool) error {
return nil
}
func (h HyperVStubber) Remove(mc *vmconfigs.MachineConfig) ([]string, func() error, error) {
// Allow removal in these two cases:
// 1. if the user is Admin
// 2. if the user has Hyper-V admin rights and there are 2+ *NEW* machines.
// *NEW* machines are those created with the vsock entry having the ToolName field.
//
// This is to prevent a non-admin user from deleting the last machine
// which would require removal of vsock entries from the Windows Registry.
if err := h.canRemove(mc); err != nil {
return nil, nil, err
}
_, vm, err := GetVMFromMC(mc)
if err != nil {
return nil, nil, err
}
rmFunc := func() error {
// Remove ignition registry entries - not a fatal error
// for vm removal
// TODO we could improve this by recommending an action be done
if err := removeIgnitionFromRegistry(mc, vm); err != nil {
logrus.Errorf("unable to remove ignition registry entries: %q", err)
}
// disk path removal is done by generic remove
if err = vm.Remove(""); err != nil {
return err
}
// remove vsock registry entries
if isLegacyMachine(mc) {
removeLegacyHvSockEntries(mc)
} else {
if err := h.removeHvSockFromRegistry(); err != nil {
logrus.Errorf("unable to remove hvsock registry entries: %q", err)
}
}
return nil
}
return []string{}, rmFunc, nil
}
func (h HyperVStubber) canCreate() error {
// if admin, can always create
if windows.HasAdminRights() {
return nil
}
// not admin, check if there is at least one existing machine
// if so, user could create more machines just by being member of the hyperv admin group
machines, err := h.countMachinesWithToolname()
if err != nil {
return err
}
// no existing machines, require to be admin
if machines == 0 {
return ErrHypervRegistryInitRequiresElevation
}
// at least 1 machine exists
// if user is member of the hyperv admin group, allow creation
if HasHyperVAdminRights() {
return nil
}
return ErrHypervUserNotInAdminGroup
}
func isLegacyMachine(mc *vmconfigs.MachineConfig) bool {
return mc.HyperVHypervisor != nil && mc.HyperVHypervisor.ReadyVsock.MachineName != ""
}
func (h HyperVStubber) canRemove(mc *vmconfigs.MachineConfig) error {
// if admin, can always remove
if windows.HasAdminRights() {
return nil
}
// if machine is legacy (machineName field), require admin rights to remove
if isLegacyMachine(mc) {
return ErrHypervRegistryRemoveRequiresElevation
}
// not admin, check if there are multiple machines
// if so, user could remove the machine just by being member of the hyperv admin group
// (only the last machine removal requires Registry changes)
machines, err := h.countMachinesWithToolname()
if err != nil {
return err
}
// more than 1 machine exists, allow removal if user has hyperv admin rights
if machines > 1 && HasHyperVAdminRights() {
return nil
}
return ErrHypervRegistryRemoveRequiresElevation
}
// canStartOrStop checks if the machine can be started or stopped.
// Legacy machines require admin rights to start or stop.
func (h HyperVStubber) canStartOrStop(mc *vmconfigs.MachineConfig) error {
if windows.HasAdminRights() {
return nil
}
// if machine is legacy (machineName field), require admin rights to start or stop
if isLegacyMachine(mc) {
return ErrHypervLegacyMachineRequiresElevation
}
return nil
}
// countMachinesWithToolname counts only machines that have a toolname field with value "podman".
func (h HyperVStubber) countMachinesWithToolname() (int, error) {
dirs, err := env.GetMachineDirs(h.VMType())
if err != nil {
return 0, err
}
mcs, err := vmconfigs.LoadMachinesInDir(dirs)
if err != nil {
return 0, err
}
count := 0
for _, mc := range mcs {
if mc.HyperVHypervisor != nil && mc.HyperVHypervisor.ReadyVsock.ToolName == "podman" {
count++
}
}
return count, nil
}
// removeLegacyHvSockEntries removes any legacy HVSOCK registry entries associated with the machine.
// This is used to clean up entries from older versions of Podman that did not manage the Toolname field.
func removeLegacyHvSockEntries(mc *vmconfigs.MachineConfig) {
if mc.HyperVHypervisor.NetworkVSock.MachineName != "" {
// Remove the HVSOCK for networking
if err := mc.HyperVHypervisor.NetworkVSock.Remove(); err != nil {
logrus.Errorf("unable to remove registry entry for %s: %q", mc.HyperVHypervisor.NetworkVSock.KeyName, err)
}
}
if mc.HyperVHypervisor.ReadyVsock.MachineName != "" {
// Remove the HVSOCK for events
if err := mc.HyperVHypervisor.ReadyVsock.Remove(); err != nil {
logrus.Errorf("unable to remove registry entry for %s: %q", mc.HyperVHypervisor.ReadyVsock.KeyName, err)
}
}
for _, mount := range mc.Mounts {
if mount.VSockNumber == nil {
// nothing to do if the vsock number was never defined
continue
}
vsockReg, err := vsock.LoadHVSockRegistryEntry(*mount.VSockNumber)
if err != nil {
logrus.Debugf("Vsock %d for mountpoint %s does not have a valid registry entry, skipping removal", *mount.VSockNumber, mount.Target)
continue
}
if vsockReg.MachineName == "" {
continue
}
if err := vsockReg.Remove(); err != nil {
logrus.Debugf("unable to remove vsock %d for mountpoint %s: %v", *mount.VSockNumber, mount.Target, err)
}
}
}
func (h HyperVStubber) removeHvSockFromRegistry() error {
// Remove hvsock registry entries only if this is the last machine
machines, err := h.countMachinesWithToolname()
if err != nil {
return err
}
if machines > 1 {
return nil
}
return vsock.RemoveAllHVSockRegistryEntries()
}
func (h HyperVStubber) RemoveAndCleanMachines(_ *define.MachineDirs) error {
return nil
}
func (h HyperVStubber) StartNetworking(mc *vmconfigs.MachineConfig, cmd *gvproxy.GvproxyCommand) error {
cmd.AddEndpoint(fmt.Sprintf("vsock://%s", mc.HyperVHypervisor.NetworkVSock.KeyName))
return nil
}
func (h HyperVStubber) StartVM(mc *vmconfigs.MachineConfig) (func() error, func() error, error) {
if err := h.canStartOrStop(mc); err != nil {
return nil, nil, err
}
_, vm, err := GetVMFromMC(mc)
if err != nil {
return nil, nil, err
}
callbackFuncs := machine.CleanUp()
defer callbackFuncs.CleanIfErr(&err)
go callbackFuncs.CleanOnSignal()
if mc.IsFirstBoot() {
// this is added because if the machine does not start
// properly on first boot, the next boot will be considered
// the first boot again and the addition of the ignition
// entries will fail because the key/value pairs already exist.
rmIgnCallbackFunc := func() error {
return removeIgnitionFromRegistry(mc, vm)
}
callbackFuncs.Add(rmIgnCallbackFunc)
// Add ignition entries to windows registry
// for first boot only
if err = readAndSplitIgnition(mc, vm); err != nil {
return nil, nil, err
}
}
waitReady, listener, err := mc.HyperVHypervisor.ReadyVsock.ListenSetupWait()
if err != nil {
return nil, nil, err
}
err = vm.Start()
if err != nil {
// cleanup the pending listener
_ = listener.Close()
return nil, nil, err
}
startCallback := func() error {
return vm.Stop()
}
callbackFuncs.Add(startCallback)
return nil, waitReady, err
}
// State is returns the state as a define.status. for hyperv, state differs from others because
// state is determined by the VM itself. normally this can be done with vm.State() and a conversion
// but doing here as well. this requires a little more interaction with the hypervisor
func (h HyperVStubber) State(mc *vmconfigs.MachineConfig, _ bool) (define.Status, error) {
_, vm, err := GetVMFromMC(mc)
if err != nil {
return define.Unknown, err
}
return stateConversion(vm.State())
}
func (h HyperVStubber) StopVM(mc *vmconfigs.MachineConfig, hardStop bool) error {
if err := h.canStartOrStop(mc); err != nil {
return err
}
vmm := hypervctl.NewVirtualMachineManager()
vm, err := vmm.GetMachine(mc.Name)
if err != nil {
return fmt.Errorf("getting virtual machine: %w", err)
}
vmState := vm.State()
if vm.State() == hypervctl.Disabled {
return nil
}
if vmState != hypervctl.Enabled { // more states could be provided as well
return hypervctl.ErrMachineStateInvalid
}
if hardStop {
return vm.StopWithForce()
}
return vm.Stop()
}
// TODO should this be plumbed higher into the code stack?
func (h HyperVStubber) StopHostNetworking(mc *vmconfigs.MachineConfig, vmType define.VMType) error {
err := machine.StopWinProxy(mc.Name, vmType)
// in podman 4, this was a "soft" error; keeping behavior as such
if err != nil {
fmt.Fprintf(os.Stderr, "Could not stop API forwarding service (win-sshproxy.exe): %s\n", err.Error())
}
return nil
}
func (h HyperVStubber) VMType() define.VMType {
return define.HyperVVirt
}
func GetVMFromMC(mc *vmconfigs.MachineConfig) (*hypervctl.VirtualMachineManager, *hypervctl.VirtualMachine, error) {
vmm := hypervctl.NewVirtualMachineManager()
vm, err := vmm.GetMachine(mc.Name)
return vmm, vm, err
}
func stateConversion(s hypervctl.EnabledState) (define.Status, error) {
switch s {
case hypervctl.Enabled:
return define.Running, nil
case hypervctl.Disabled:
return define.Stopped, nil
case hypervctl.Starting:
return define.Starting, nil
}
return define.Unknown, fmt.Errorf("unknown state: %q", s.String())
}
func (h HyperVStubber) SetProviderAttrs(mc *vmconfigs.MachineConfig, opts define.SetOptions) error {
var cpuChanged, memoryChanged bool
_, vm, err := GetVMFromMC(mc)
if err != nil {
return err
}
if vm.State() != hypervctl.Disabled {
return errors.New("unable to change settings unless vm is stopped")
}
if opts.Rootful != nil && mc.HostUser.Rootful != *opts.Rootful {
if err := mc.SetRootful(*opts.Rootful); err != nil {
return err
}
}
if opts.DiskSize != nil {
if err := resizeDisk(*opts.DiskSize, mc.ImagePath); err != nil {
return err
}
}
if opts.CPUs != nil {
cpuChanged = true
}
if opts.Memory != nil {
memoryChanged = true
}
if cpuChanged || memoryChanged {
err := vm.UpdateProcessorMemSettings(func(ps *hypervctl.ProcessorSettings) {
if cpuChanged {
ps.VirtualQuantity = *opts.CPUs
}
}, func(ms *hypervctl.MemorySettings) {
if memoryChanged {
ms.DynamicMemoryEnabled = false
mem := uint64(*opts.Memory)
ms.VirtualQuantity = mem
ms.Limit = mem
ms.Reservation = mem
}
})
if err != nil {
return fmt.Errorf("setting CPU and Memory for VM: %w", err)
}
}
if opts.USBs != nil {
return fmt.Errorf("changing USBs not supported for hyperv machines")
}
return nil
}
func (h HyperVStubber) PrepareIgnition(mc *vmconfigs.MachineConfig, _ *ignition.IgnitionBuilder) (*ignition.ReadyUnitOpts, error) {
// HyperV is different because it has to know some ignition details before creating the VM. It cannot
// simply be derived. So we create the HyperVConfig here.
mc.HyperVHypervisor = new(vmconfigs.HyperVConfig)
var ignOpts ignition.ReadyUnitOpts
// Attempt to load an existing HVSock registry entry for events.
// If no existing entry is found, create a new one.
// Creating a new entry requires administrative rights.
readySock, err := vsock.LoadHVSockRegistryEntryByPurpose(vsock.Events)
if err != nil {
if !windows.HasAdminRights() {
return nil, ErrHypervRegistryInitRequiresElevation
}
readySock, err = vsock.NewHVSockRegistryEntry(vsock.Events)
if err != nil {
return nil, err
}
}
// TODO Stopped here ... fails bc mc.Hypervisor is nil ... this can be nil checked prior and created
// however the same will have to be done in create
mc.HyperVHypervisor.ReadyVsock = *readySock
ignOpts.Port = readySock.Port
return &ignOpts, nil
}
func (h HyperVStubber) PostStartNetworking(mc *vmconfigs.MachineConfig, _ bool) error {
var (
err error
executable string
)
callbackFuncs := machine.CleanUp()
defer callbackFuncs.CleanIfErr(&err)
go callbackFuncs.CleanOnSignal()
if len(mc.Mounts) == 0 {
return nil
}
var (
dirs *define.MachineDirs
gvproxyPID int
)
dirs, err = env.GetMachineDirs(h.VMType())
if err != nil {
return err
}
// GvProxy PID file path is now derived
gvproxyPIDFile, err := dirs.RuntimeDir.AppendToNewVMFile("gvproxy.pid", nil)
if err != nil {
return err
}
gvproxyPID, err = gvproxyPIDFile.ReadPIDFrom()
if err != nil {
return err
}
executable, err = os.Executable()
if err != nil {
return err
}
// Start the 9p server in the background
p9ServerArgs := []string{}
if logrus.IsLevelEnabled(logrus.DebugLevel) {
p9ServerArgs = append(p9ServerArgs, "--log-level=debug")
}
p9ServerArgs = append(p9ServerArgs, "machine", "server9p")
for _, mount := range mc.Mounts {
if mount.VSockNumber == nil {
return fmt.Errorf("mount %s has no vsock port defined", mount.Source)
}
p9ServerArgs = append(p9ServerArgs, "--serve", fmt.Sprintf("%s:%s", mount.Source, winio.VsockServiceID(uint32(*mount.VSockNumber)).String()))
}
p9ServerArgs = append(p9ServerArgs, fmt.Sprintf("%d", gvproxyPID))
logrus.Debugf("Going to start 9p server using command: %s %v", executable, p9ServerArgs)
fsCmd := exec.Command(executable, p9ServerArgs...)
if logrus.IsLevelEnabled(logrus.DebugLevel) {
log, err := logCommandToFile(fsCmd, "podman-machine-server9.log")
if err != nil {
return err
}
defer log.Close()
}
err = fsCmd.Start()
if err != nil {
return fmt.Errorf("unable to start 9p server: %v", err)
}
logrus.Infof("Started podman 9p server as PID %d", fsCmd.Process.Pid)
// Note: No callback is needed to stop the 9p server, because it will stop when
// gvproxy stops
// Finalize starting shares after we are confident gvproxy is still alive.
err = startShares(mc)
return err
}
func (h HyperVStubber) UpdateSSHPort(_ *vmconfigs.MachineConfig, _ int) error {
// managed by gvproxy on this backend, so nothing to do
return nil
}
func resizeDisk(newSize strongunits.GiB, imagePath *define.VMFile) error {
resize := exec.Command("powershell", []string{"-command", fmt.Sprintf("Resize-VHD \"%s\" %d", imagePath.GetPath(), newSize.ToBytes())}...)
logrus.Debug(resize.Args)
resize.Stdout = os.Stdout
resize.Stderr = os.Stderr
if err := resize.Run(); err != nil {
return fmt.Errorf("resizing image: %q", err)
}
return nil
}
func getIgnitionReader(mc *vmconfigs.MachineConfig) (*bytes.Reader, error) {
ignFile, err := mc.IgnitionFile()
if err != nil {
return nil, err
}
ign, err := ignFile.Read()
if err != nil {
return nil, err
}
return bytes.NewReader(ign), nil
}
// readAndSplitIgnition reads the ignition file and splits it into key:value pairs
func readAndSplitIgnition(mc *vmconfigs.MachineConfig, vm *hypervctl.VirtualMachine) error {
reader, err := getIgnitionReader(mc)
if err != nil {
return err
}
return vm.SplitAndAddIgnition("ignition.config.", reader)
}
func removeIgnitionFromRegistry(mc *vmconfigs.MachineConfig, vm *hypervctl.VirtualMachine) error {
// because the vm is down at this point, we cannot query hyperv for these key value pairs.
// therefore we recalculate the number of parts ignition file consists of and
// remove all ignition key/value pairs the same way they were added
reader, err := getIgnitionReader(mc)
if err != nil {
return err
}
parts, err := ginsu.Dice(reader)
if err != nil {
return err
}
for idx := range parts {
key := fmt.Sprintf("ignition.config.%d", idx)
if err := vm.RemoveKeyValuePairNoWait(key); err != nil {
return err
}
}
return nil
}
func logCommandToFile(c *exec.Cmd, filename string) (*os.File, error) {
dir, err := env.GetDataDir(define.HyperVVirt)
if err != nil {
return nil, fmt.Errorf("obtain machine dir: %w", err)
}
path := filepath.Join(dir, filename)
logrus.Infof("Going to log to %s", path)
log, err := os.Create(path)
if err != nil {
return nil, fmt.Errorf("create log file: %w", err)
}
c.Stdout = log
c.Stderr = log
return log, nil
}
const hyperVVsockNMConnection = `
[connection]
id=vsock0
type=tun
interface-name=vsock0
[tun]
mode=2
[802-3-ethernet]
cloned-mac-address=5A:94:EF:E4:0C:EE
[ipv4]
method=auto
[proxy]
`
func createNetworkUnit(netPort uint64) (string, error) {
netUnit := parser.NewUnitFile()
netUnit.Add("Unit", "Description", "vsock_network")
netUnit.Add("Unit", "After", "NetworkManager.service")
netUnit.Add("Service", "ExecStart", fmt.Sprintf("/usr/libexec/podman/gvforwarder -preexisting -iface vsock0 -url vsock://2:%d/connect", netPort))
netUnit.Add("Service", "ExecStartPost", "/usr/bin/nmcli c up vsock0")
netUnit.Add("Install", "WantedBy", "multi-user.target")
return netUnit.ToString()
}
func (h HyperVStubber) GetRosetta(_ *vmconfigs.MachineConfig) (bool, error) {
return false, nil
}