banger/internal/imagepull/inject.go
Thales Maciel c3fb4ccc3e
Phase B-3: first-boot sshd install
New internal/imagepull/assets/first-boot.sh: POSIX-sh oneshot that
detects the guest distro from /etc/os-release (ID + ID_LIKE
fallback), installs openssh-server via the native package manager,
and enables/starts sshd. Covers debian/ubuntu/kali/raspbian/pop,
alpine, fedora/rhel/centos/rocky/almalinux, arch/manjaro, and
opensuse/suse. Unknown distros fail clearly with a pointer at
editing the script to add a branch.

Marker-driven: the service has ConditionPathExists=
/var/lib/banger/first-boot-pending, and the script removes the
marker on success. Subsequent boots no-op.

Testability seams in the script: RUN_PLAN=1 skips the
sshd-already-present short-circuit and makes the dispatch echo the
planned command instead of executing it. OS_RELEASE_FILE and
BANGER_FIRST_BOOT_MARKER env vars override paths so the Go tests
exercise the real dispatch logic in a tempdir without touching
/etc or /var/lib on the host.

Embedding: internal/imagepull/firstboot.go go:embeds both the
script and the systemd unit; exposes FirstBootScript() and
FirstBootUnit() plus the FirstBootScriptPath /
FirstBootMarkerPath / FirstBootUnitName constants.

Injection: InjectGuestAgents now drops /usr/local/libexec/
banger-first-boot (0755), /etc/systemd/system/banger-first-boot.
service (0644), the empty /var/lib/banger/first-boot-pending
marker (0644), and the multi-user.target.wants enable symlink.
All uid=0, gid=0.

Tests: eight-case dispatch-by-distro (debian, ubuntu, alpine,
fedora, arch, opensuse, plus ID_LIKE fallbacks for weird
derivatives). Script syntax check via `sh -n`. Unit-contains-
expected-fields check. Existing inject round-trip test extended
to assert the first-boot bits land in the ext4.

Deferred: per-image FirstBootPending flag + extended SSH wait
timeout at VM start. Will add if live verification (B-4) shows
the naive retry UX is unacceptable.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-04-16 18:20:33 -03:00

248 lines
7.4 KiB
Go

package imagepull
import (
"bytes"
"context"
"fmt"
"os"
"path/filepath"
"sort"
"strings"
"banger/internal/guestnet"
"banger/internal/system"
"banger/internal/vsockagent"
)
// GuestAgentAssets bundles everything the guest side of banger needs in a
// rootfs that doesn't already have it. Callers (the daemon's PullImage)
// resolve the vsock-agent binary path via paths.CompanionBinaryPath and
// hand it in; the rest comes from the respective asset packages.
type GuestAgentAssets struct {
VsockAgentBin string // absolute path on the host, copied verbatim
}
// InjectGuestAgents writes banger's guest-side assets (vsock agent
// binary + systemd unit, network bootstrap script + unit, vsock modules-
// load config, symlinks that enable the units at boot) into ext4File.
// All entries land with uid=0, gid=0 and appropriate modes.
//
// Runs in one debugfs -w invocation: dirs, files, sif (uid/gid/mode),
// and symlinks all in one scripted batch. No sudo required because the
// ext4 is owned by the runner.
func InjectGuestAgents(ctx context.Context, runner system.CommandRunner, ext4File string, assets GuestAgentAssets) error {
if assets.VsockAgentBin == "" {
return fmt.Errorf("vsock-agent binary path is required")
}
if _, err := os.Stat(assets.VsockAgentBin); err != nil {
return fmt.Errorf("vsock-agent binary %q missing: %w", assets.VsockAgentBin, err)
}
// Stage content blobs as temp files so debugfs `write` can pick
// them up. All other commands (mkdir/sif/symlink) are inline.
stage, err := os.MkdirTemp("", "banger-inject-")
if err != nil {
return err
}
defer os.RemoveAll(stage)
steps := []injectFile{
{
hostSrc: assets.VsockAgentBin,
guestPath: vsockagent.GuestInstallPath, // /usr/local/bin/banger-vsock-agent
mode: 0o755,
},
{
content: []byte(guestnet.BootstrapScript()),
guestPath: guestnet.GuestScriptPath, // /usr/local/libexec/banger-network-bootstrap
mode: 0o755,
},
{
content: []byte(guestnet.SystemdServiceUnit()),
guestPath: "/etc/systemd/system/" + guestnet.SystemdServiceName, // banger-network.service
mode: 0o644,
},
{
content: []byte(vsockagent.ServiceUnit()),
guestPath: "/etc/systemd/system/" + vsockagent.ServiceName, // banger-vsock-agent.service
mode: 0o644,
},
{
content: []byte(vsockagent.ModulesLoadConfig()),
guestPath: "/etc/modules-load.d/banger-vsock.conf",
mode: 0o644,
},
{
content: []byte(FirstBootScript()),
guestPath: FirstBootScriptPath, // /usr/local/libexec/banger-first-boot
mode: 0o755,
},
{
content: []byte(FirstBootUnit()),
guestPath: "/etc/systemd/system/" + FirstBootUnitName,
mode: 0o644,
},
{
content: nil, // empty marker file — its existence triggers the service
guestPath: FirstBootMarkerPath,
mode: 0o644,
},
}
// Resolve content-backed steps to on-disk temp files.
for i := range steps {
if steps[i].hostSrc != "" {
continue
}
tmp := filepath.Join(stage, fmt.Sprintf("blob-%d", i))
if err := os.WriteFile(tmp, steps[i].content, 0o644); err != nil {
return err
}
steps[i].hostSrc = tmp
}
symlinks := []injectSymlink{
{
target: "/etc/systemd/system/" + guestnet.SystemdServiceName,
link: "/etc/systemd/system/multi-user.target.wants/" + guestnet.SystemdServiceName,
},
{
target: "/etc/systemd/system/" + vsockagent.ServiceName,
link: "/etc/systemd/system/multi-user.target.wants/" + vsockagent.ServiceName,
},
{
target: "/etc/systemd/system/" + FirstBootUnitName,
link: "/etc/systemd/system/multi-user.target.wants/" + FirstBootUnitName,
},
}
script := buildInjectScript(steps, symlinks)
stdinRunner, ok := runner.(system.StdinRunner)
if !ok {
return fmt.Errorf("inject requires a runner that supports stdin (got %T)", runner)
}
out, err := stdinRunner.RunStdin(ctx, script, "debugfs", "-w", "-f", "-", ext4File)
if err != nil {
return fmt.Errorf("debugfs inject: %w: %s", err, string(out))
}
// Scan output for hard errors — debugfs keeps going past errors
// with -f, so we need to look at stdout/stderr-as-stdout for bad
// signs. mkdir errors on already-present dirs are expected; we
// ignore "File exists" and "Is a directory". Other errors bubble.
if bad := scanInjectOutput(out); bad != "" {
return fmt.Errorf("debugfs inject: %s", bad)
}
return nil
}
type injectFile struct {
content []byte
hostSrc string // set by InjectGuestAgents after staging
guestPath string
mode uint32 // perm bits; type bits added by buildInjectScript
}
type injectSymlink struct {
target string
link string
}
// buildInjectScript emits the debugfs command stream.
func buildInjectScript(files []injectFile, symlinks []injectSymlink) *bytes.Buffer {
var buf bytes.Buffer
// Create every ancestor directory of every file/symlink path. mkdir
// on an already-existing dir is benign (debugfs continues past the
// error), but we prune duplicates to keep the script clean.
dirs := collectAncestors(files, symlinks)
for _, d := range dirs {
fmt.Fprintf(&buf, "mkdir %s\n", d)
}
// Write each file content.
for _, f := range files {
fmt.Fprintf(&buf, "write %s %s\n", f.hostSrc, f.guestPath)
}
// Fix ownership + mode on every written file (uid=0, gid=0).
for _, f := range files {
fmt.Fprintf(&buf, "set_inode_field %s uid 0\n", f.guestPath)
fmt.Fprintf(&buf, "set_inode_field %s gid 0\n", f.guestPath)
fmt.Fprintf(&buf, "set_inode_field %s mode 0%o\n", f.guestPath, 0o100000|f.mode)
}
// Fix dir ownership. Don't touch modes — mkdir's default 0755 is fine.
for _, d := range dirs {
fmt.Fprintf(&buf, "set_inode_field %s uid 0\n", d)
fmt.Fprintf(&buf, "set_inode_field %s gid 0\n", d)
}
// Finally, create the enable-at-boot symlinks.
for _, s := range symlinks {
fmt.Fprintf(&buf, "symlink %s %s\n", s.link, s.target)
}
return &buf
}
// collectAncestors walks every file + symlink path and returns the unique
// set of parent directories, sorted shallowest first so mkdir ordering
// is valid.
func collectAncestors(files []injectFile, symlinks []injectSymlink) []string {
set := map[string]struct{}{}
add := func(p string) {
dir := filepath.Dir(p)
for dir != "" && dir != "/" {
set[dir] = struct{}{}
dir = filepath.Dir(dir)
}
}
for _, f := range files {
add(f.guestPath)
}
for _, s := range symlinks {
add(s.link)
}
out := make([]string, 0, len(set))
for d := range set {
out = append(out, d)
}
// Shallow-first by depth, then lexicographic.
sort.Slice(out, func(i, j int) bool {
di := strings.Count(out[i], "/")
dj := strings.Count(out[j], "/")
if di != dj {
return di < dj
}
return out[i] < out[j]
})
return out
}
// scanInjectOutput returns a non-empty string if debugfs reported an
// error that's not a benign "File exists" from mkdir on an already-
// present directory. Debugfs emits errors on stderr AND stdout (which
// we capture together); we look for known failure signatures.
func scanInjectOutput(out []byte) string {
lines := strings.Split(string(out), "\n")
for _, line := range lines {
line = strings.TrimSpace(line)
if line == "" {
continue
}
// Benign: mkdir on existing dir.
if strings.Contains(line, "File exists") {
continue
}
// Failure signatures we care about.
if strings.Contains(line, "error writing file") ||
strings.Contains(line, "couldn't find") ||
strings.Contains(line, "No such file") ||
strings.Contains(line, "Unrecognized command") ||
strings.Contains(line, "symlink:") {
return line
}
}
return ""
}