mcp/internal/runtime/qemu.go

package runtime

import (
	"context"
	"crypto/sha256"
	"encoding/hex"
	"encoding/json"
	"fmt"
	"net"
	"os"
	"os/exec"
	"path/filepath"
	"strconv"
	"strings"
	"syscall"
	"time"
)

// dialUnix connects to a unix-domain socket with a timeout.
func dialUnix(path string, timeout time.Duration) (net.Conn, error) {
	return net.DialTimeout("unix", path, timeout)
}

// QEMU implements the Runtime interface by running services as Nanos
// unikernel virtual machines under QEMU/KVM instead of containers.
//
// Each service component becomes a single-process VM with its own kernel.
// The lifecycle maps onto the same Runtime interface as Podman so the agent
// can treat unikernels and containers uniformly:
//
//	Pull    -> pull the OCI image, extract the ELF binary, cache it
//	Run     -> `ops build` the binary into a Nanos image, boot it under QEMU
//	Stop    -> graceful QMP powerdown, then SIGTERM/SIGKILL the QEMU process
//	Remove  -> stop and delete the VM state directory
//	Inspect -> read persisted metadata + check process liveness
//	List    -> enumerate VM state directories
//	Logs    -> stream the serial console log file
//
// Phase 1 uses QEMU user-mode networking with host port forwards, which is
// functionally equivalent to rootless podman's localhost port mappings:
// mc-proxy routes to 127.0.0.1:<hostport> exactly as it does for containers.
// Isolated bridge networking is a later phase.
type QEMU struct {
	// ImageDir holds built Nanos images and extracted binaries.
	// Default: /srv/mcp/images
	ImageDir string
	// StateDir holds per-VM runtime state (pidfile, QMP socket, console log,
	// metadata). Default: /srv/mcp/vm
	StateDir string
	// OpsPath is the path to the `ops` Nanos toolchain binary. Default: "ops".
	OpsPath string
	// QemuPath is the path to qemu-system-x86_64. Default: "qemu-system-x86_64".
	QemuPath string
	// Memory is the default guest memory in MB when a spec does not set one.
	Memory int
	// HomeDir is set as $HOME for `ops` so it uses a stable ~/.ops directory.
	HomeDir string
}

func (q *QEMU) imageDir() string {
	if q.ImageDir != "" {
		return q.ImageDir
	}
	return "/srv/mcp/images"
}

func (q *QEMU) stateDir() string {
	if q.StateDir != "" {
		return q.StateDir
	}
	return "/srv/mcp/vm"
}

func (q *QEMU) opsPath() string {
	if q.OpsPath != "" {
		return q.OpsPath
	}
	return "ops"
}

func (q *QEMU) qemuPath() string {
	if q.QemuPath != "" {
		return q.QemuPath
	}
	return "qemu-system-x86_64"
}

func (q *QEMU) memory() int {
	if q.Memory > 0 {
		return q.Memory
	}
	return 256
}

// opsEnv returns the environment for invoking `ops`, pinning $HOME so its
// cache and image directory are stable across invocations.
func (q *QEMU) opsEnv() []string {
	env := os.Environ()
	if q.HomeDir != "" {
		env = append(env, "HOME="+q.HomeDir)
	}
	return env
}

// sanitizeImage turns an image reference into a filesystem-safe stem.
//
//	"mcr.example:8443/mcdoc:v0.1.0" -> "mcr.example_8443_mcdoc_v0.1.0"
func sanitizeImage(image string) string {
	r := strings.NewReplacer("/", "_", ":", "_")
	return r.Replace(image)
}

// binaryName derives the in-image ELF binary name from an image reference by
// taking the repository basename. "host:8443/mcdoc:v0.1.0" -> "mcdoc".
func binaryName(image string) string {
	name := image
	if i := strings.LastIndex(name, "/"); i >= 0 {
		name = name[i+1:]
	}
	if i := strings.Index(name, ":"); i >= 0 {
		name = name[:i]
	}
	return name
}

func (q *QEMU) binPath(image string) string {
	return filepath.Join(q.imageDir(), sanitizeImage(image)+".bin")
}

func (q *QEMU) imgPath(name string) string {
	return filepath.Join(q.imageDir(), name+".img")
}

func (q *QEMU) vmDir(name string) string {
	return filepath.Join(q.stateDir(), name)
}

// vmMeta is the persisted per-VM metadata written at Run time so that
// Inspect/List can report accurate information after an agent restart.
type vmMeta struct {
	Name      string    `json:"name"`
	Image     string    `json:"image"`
	User      string    `json:"user"`
	Restart   string    `json:"restart"`
	Ports     []string  `json:"ports"`
	Volumes   []string  `json:"volumes"`
	Cmd       []string  `json:"cmd"`
	MemoryMB  int       `json:"memory_mb"`
	VCPUs     int       `json:"vcpus"`
	ImageHash string    `json:"image_hash"`
	Started   time.Time `json:"started"`
}

// Pull pulls the OCI image and extracts its ELF binary into the image cache.
// The binary is the input to `ops build`; the Nanos image itself is built at
// Run time so the service's command arguments can be baked in.
func (q *QEMU) Pull(ctx context.Context, image string) error {
	if err := os.MkdirAll(q.imageDir(), 0o750); err != nil {
		return fmt.Errorf("create image dir: %w", err)
	}

	// Pull the OCI image via podman (reuses the agent's registry auth).
	if out, err := exec.CommandContext(ctx, "podman", "pull", image).CombinedOutput(); err != nil { //nolint:gosec // args built programmatically
		return fmt.Errorf("podman pull %q: %w: %s", image, err, out)
	}

	// Create (do not start) a container to copy the binary out of.
	tmp := "ukextract-" + sanitizeImage(image)
	_ = exec.CommandContext(ctx, "podman", "rm", "-f", tmp).Run()                                                    //nolint:gosec
	if out, err := exec.CommandContext(ctx, "podman", "create", "--name", tmp, image).CombinedOutput(); err != nil { //nolint:gosec
		return fmt.Errorf("podman create %q: %w: %s", image, err, out)
	}
	defer func() { _ = exec.Command("podman", "rm", "-f", tmp).Run() }() //nolint:gosec

	bin := binaryName(image)
	src := tmp + ":/usr/local/bin/" + bin
	dst := q.binPath(image)
	if out, err := exec.CommandContext(ctx, "podman", "cp", src, dst).CombinedOutput(); err != nil { //nolint:gosec
		return fmt.Errorf("extract binary %q from %q: %w: %s", bin, image, err, out)
	}
	if err := os.Chmod(dst, 0o755); err != nil { //nolint:gosec // unikernel ELF must be executable
		return fmt.Errorf("chmod extracted binary: %w", err)
	}
	return nil
}

// opsConfig is the subset of the `ops` build configuration we generate.
type opsConfig struct {
	Args      []string          `json:"Args,omitempty"`
	Env       map[string]string `json:"Env,omitempty"`
	RunConfig opsRunConfig      `json:"RunConfig"`
}

type opsRunConfig struct {
	Ports    []string `json:"Ports,omitempty"`
	Memory   string   `json:"Memory,omitempty"`
	CPUs     int      `json:"CPUs,omitempty"`
	Klibs    []string `json:"Klibs,omitempty"`
	Mounts   any      `json:"Mounts,omitempty"`
	NoTrace  []string `json:"NoTrace,omitempty"`
	GDBPort  int      `json:"GDBPort,omitempty"`
	Nanos    string   `json:"Nanos,omitempty"`
	Hostname string   `json:"Hostname,omitempty"`
}

// guestPorts extracts the guest (container) port from each spec port mapping.
// Accepts "host:container", "ip:host:container", or a bare "port".
func guestPorts(ports []string) []string {
	var gp []string
	for _, p := range ports {
		parts := strings.Split(p, ":")
		gp = append(gp, parts[len(parts)-1])
	}
	return gp
}

// hostForward builds the QEMU hostfwd value for a spec port mapping.
// "ip:host:container" -> "tcp:ip:host-:container"
// "host:container"    -> "tcp:127.0.0.1:host-:container"
func hostForward(p string) string {
	parts := strings.Split(p, ":")
	switch len(parts) {
	case 3:
		return fmt.Sprintf("tcp:%s:%s-:%s", parts[0], parts[1], parts[2])
	case 2:
		return fmt.Sprintf("tcp:127.0.0.1:%s-:%s", parts[0], parts[1])
	default:
		return fmt.Sprintf("tcp:127.0.0.1:%s-:%s", parts[0], parts[0])
	}
}

// Run builds the Nanos image (if needed) and boots it under QEMU/KVM.
func (q *QEMU) Run(ctx context.Context, spec ContainerSpec) error {
	if err := os.MkdirAll(q.vmDir(spec.Name), 0o750); err != nil {
		return fmt.Errorf("create vm state dir: %w", err)
	}

	mem := spec.MemoryMB
	if mem <= 0 {
		mem = q.memory()
	}
	cpus := spec.VCPUs
	if cpus <= 0 {
		cpus = 1
	}

	// Build the Nanos image from the extracted binary, baking in command args.
	cfg := opsConfig{
		Args: spec.Cmd,
		RunConfig: opsRunConfig{
			Ports:  guestPorts(spec.Ports),
			Memory: strconv.Itoa(mem) + "m",
			CPUs:   cpus,
		},
	}
	if len(spec.Env) > 0 {
		cfg.Env = map[string]string{}
		for _, e := range spec.Env {
			if i := strings.Index(e, "="); i >= 0 {
				cfg.Env[e[:i]] = e[i+1:]
			}
		}
	}
	cfgPath := filepath.Join(q.vmDir(spec.Name), "ops.json")
	cfgBytes, err := json.MarshalIndent(cfg, "", "  ")
	if err != nil {
		return fmt.Errorf("marshal ops config: %w", err)
	}
	if err := os.WriteFile(cfgPath, cfgBytes, 0o640); err != nil { //nolint:gosec // mcp-group-readable config
		return fmt.Errorf("write ops config: %w", err)
	}

	img := q.imgPath(spec.Name)
	bin := q.binPath(spec.Image)
	if _, err := os.Stat(bin); err != nil {
		return fmt.Errorf("binary not found for %q (Pull first): %w", spec.Image, err)
	}
	build := exec.CommandContext(ctx, q.opsPath(), "build", bin, "-c", cfgPath, "-i", spec.Name) //nolint:gosec
	build.Env = q.opsEnv()
	if out, err := build.CombinedOutput(); err != nil {
		return fmt.Errorf("ops build %q: %w: %s", spec.Name, err, out)
	}
	// ops writes to ~/.ops/images/<name>.img; move it into our image dir.
	opsImg := filepath.Join(q.opsImagesDir(), spec.Name+".img")
	if _, err := os.Stat(opsImg); err == nil {
		if err := os.Rename(opsImg, img); err != nil {
			// Cross-device fallback: copy.
			if cpErr := copyFile(opsImg, img); cpErr != nil {
				return fmt.Errorf("relocate built image: %w", err)
			}
			_ = os.Remove(opsImg)
		}
	}

	hash, _ := fileSHA256(img)

	// Assemble QEMU invocation: KVM-accelerated, headless, serial console to a
	// file, QMP control socket, virtio disk + NIC with user-mode port forwards.
	netdev := "user,id=n0"
	for _, p := range spec.Ports {
		netdev += ",hostfwd=" + hostForward(p)
	}
	args := []string{
		"-enable-kvm",
		"-m", strconv.Itoa(mem),
		"-smp", strconv.Itoa(cpus),
		"-display", "none",
		"-no-reboot",
		"-daemonize",
		"-pidfile", filepath.Join(q.vmDir(spec.Name), "qemu.pid"),
		"-serial", "file:" + filepath.Join(q.vmDir(spec.Name), "console.log"),
		"-qmp", "unix:" + filepath.Join(q.vmDir(spec.Name), "qmp.sock") + ",server,nowait",
		"-drive", "file=" + img + ",format=raw,if=virtio",
		"-device", "virtio-net-pci,netdev=n0",
		"-netdev", netdev,
	}
	// 9p passthrough for host /srv/<service> volumes (best-effort; Nanos must
	// support the 9p client for the guest to mount it).
	for i, v := range spec.Volumes {
		parts := strings.SplitN(v, ":", 2)
		host := parts[0]
		tag := fmt.Sprintf("srv%d", i)
		args = append(args,
			"-virtfs", fmt.Sprintf("local,path=%s,mount_tag=%s,security_model=none,id=%s", host, tag, tag),
		)
	}

	cmd := exec.CommandContext(ctx, q.qemuPath(), args...) //nolint:gosec
	if out, err := cmd.CombinedOutput(); err != nil {
		return fmt.Errorf("qemu launch %q: %w: %s", spec.Name, err, out)
	}

	meta := vmMeta{
		Name:      spec.Name,
		Image:     spec.Image,
		User:      spec.User,
		Restart:   spec.Restart,
		Ports:     spec.Ports,
		Volumes:   spec.Volumes,
		Cmd:       spec.Cmd,
		MemoryMB:  mem,
		VCPUs:     cpus,
		ImageHash: hash,
		Started:   time.Now().UTC(),
	}
	return q.writeMeta(spec.Name, meta)
}

func (q *QEMU) opsImagesDir() string {
	home := q.HomeDir
	if home == "" {
		home, _ = os.UserHomeDir()
	}
	return filepath.Join(home, ".ops", "images")
}

func (q *QEMU) writeMeta(name string, m vmMeta) error {
	b, err := json.MarshalIndent(m, "", "  ")
	if err != nil {
		return fmt.Errorf("marshal vm meta: %w", err)
	}
	return os.WriteFile(filepath.Join(q.vmDir(name), "meta.json"), b, 0o640) //nolint:gosec // mcp-group-readable metadata
}

func (q *QEMU) readMeta(name string) (vmMeta, error) {
	var m vmMeta
	b, err := os.ReadFile(filepath.Join(q.vmDir(name), "meta.json"))
	if err != nil {
		return m, err
	}
	return m, json.Unmarshal(b, &m)
}

// pidOf returns the running QEMU pid for a VM, or 0 if not running.
func (q *QEMU) pidOf(name string) int {
	b, err := os.ReadFile(filepath.Join(q.vmDir(name), "qemu.pid"))
	if err != nil {
		return 0
	}
	pid, err := strconv.Atoi(strings.TrimSpace(string(b)))
	if err != nil || pid <= 0 {
		return 0
	}
	if err := syscall.Kill(pid, 0); err != nil {
		return 0
	}
	return pid
}

// Stop gracefully powers down the VM, falling back to SIGTERM/SIGKILL.
func (q *QEMU) Stop(ctx context.Context, name string) error {
	pid := q.pidOf(name)
	if pid == 0 {
		return nil
	}
	// Try a graceful QMP system_powerdown.
	_ = q.qmpCommand(name, "system_powerdown")
	deadline := time.Now().Add(10 * time.Second)
	for time.Now().Before(deadline) {
		if q.pidOf(name) == 0 {
			return nil
		}
		time.Sleep(300 * time.Millisecond)
	}
	// Escalate.
	_ = syscall.Kill(pid, syscall.SIGTERM)
	time.Sleep(2 * time.Second)
	if q.pidOf(name) != 0 {
		_ = syscall.Kill(pid, syscall.SIGKILL)
	}
	return nil
}

// qmpCommand sends a single QMP command over the VM's control socket.
func (q *QEMU) qmpCommand(name, command string) error {
	sock := filepath.Join(q.vmDir(name), "qmp.sock")
	conn, err := dialUnix(sock, 3*time.Second)
	if err != nil {
		return err
	}
	defer func() { _ = conn.Close() }()
	// QMP handshake: read greeting, send qmp_capabilities, then the command.
	dec := json.NewDecoder(conn)
	var greeting map[string]any
	_ = dec.Decode(&greeting)
	_, _ = conn.Write([]byte(`{"execute":"qmp_capabilities"}`))
	var ack map[string]any
	_ = dec.Decode(&ack)
	_, err = conn.Write([]byte(`{"execute":"` + command + `"}`))
	return err
}

// Remove stops the VM and deletes its state directory.
func (q *QEMU) Remove(ctx context.Context, name string) error {
	_ = q.Stop(ctx, name)
	if pid := q.pidOf(name); pid != 0 {
		_ = syscall.Kill(pid, syscall.SIGKILL)
	}
	return os.RemoveAll(q.vmDir(name))
}

// Inspect reports the observed state of a VM.
func (q *QEMU) Inspect(ctx context.Context, name string) (ContainerInfo, error) {
	m, err := q.readMeta(name)
	if err != nil {
		return ContainerInfo{}, fmt.Errorf("qemu inspect %q: %w", name, err)
	}
	state := "stopped"
	if q.pidOf(name) != 0 {
		state = "running"
	}
	return q.infoFromMeta(m, state), nil
}

func (q *QEMU) infoFromMeta(m vmMeta, state string) ContainerInfo {
	return ContainerInfo{
		Name:    m.Name,
		Image:   m.Image,
		State:   state,
		Network: "user",
		User:    m.User,
		Restart: m.Restart,
		Ports:   m.Ports,
		Volumes: m.Volumes,
		Cmd:     m.Cmd,
		Version: ExtractVersion(m.Image),
		Started: m.Started,
	}
}

// List enumerates all VMs known from the state directory.
func (q *QEMU) List(ctx context.Context) ([]ContainerInfo, error) {
	entries, err := os.ReadDir(q.stateDir())
	if err != nil {
		if os.IsNotExist(err) {
			return nil, nil
		}
		return nil, fmt.Errorf("read vm state dir: %w", err)
	}
	var infos []ContainerInfo
	for _, e := range entries {
		if !e.IsDir() {
			continue
		}
		m, err := q.readMeta(e.Name())
		if err != nil {
			continue
		}
		state := "stopped"
		if q.pidOf(e.Name()) != 0 {
			state = "running"
		}
		infos = append(infos, q.infoFromMeta(m, state))
	}
	return infos, nil
}

// Logs streams the VM's serial console log.
func (q *QEMU) Logs(ctx context.Context, name string, tail int, follow, timestamps bool, since string) *exec.Cmd {
	console := filepath.Join(q.vmDir(name), "console.log")
	args := []string{}
	if follow {
		args = append(args, "-f")
	}
	if tail > 0 {
		args = append(args, "-n", strconv.Itoa(tail))
	} else {
		args = append(args, "-n", "+1")
	}
	args = append(args, console)
	return exec.CommandContext(ctx, "tail", args...) //nolint:gosec
}

// Build builds a Nanos image from a context directory's binary. Used by the
// `mcp build --unikernel` path. Not the primary deploy path.
func (q *QEMU) Build(ctx context.Context, image, contextDir, dockerfile string) error {
	return fmt.Errorf("qemu build: not implemented; build OCI image then Pull")
}

// Push is not implemented for the QEMU runtime.
func (q *QEMU) Push(ctx context.Context, image string) error {
	return fmt.Errorf("qemu push: not implemented")
}

// ImageExists reports whether the extracted binary for the image is cached.
func (q *QEMU) ImageExists(ctx context.Context, image string) (bool, error) {
	if _, err := os.Stat(q.binPath(image)); err == nil {
		return true, nil
	}
	return false, nil
}

// Login delegates registry auth to podman (shared credential store).
func (q *QEMU) Login(ctx context.Context, registry, username, token string) error {
	cmd := exec.CommandContext(ctx, "podman", "login", "--username", username, "--password-stdin", registry) //nolint:gosec
	cmd.Stdin = strings.NewReader(token)
	if out, err := cmd.CombinedOutput(); err != nil {
		return fmt.Errorf("podman login %q: %w: %s", registry, err, out)
	}
	return nil
}

func fileSHA256(path string) (string, error) {
	b, err := os.ReadFile(path) //nolint:gosec // hashing a known image path
	if err != nil {
		return "", err
	}
	sum := sha256.Sum256(b)
	return hex.EncodeToString(sum[:]), nil
}

func copyFile(src, dst string) error {
	in, err := os.ReadFile(src) //nolint:gosec // relocating a built image
	if err != nil {
		return err
	}
	return os.WriteFile(dst, in, 0o640) //nolint:gosec // relocating a built image
}