| 1 | // SPDX-License-Identifier: AGPL-3.0-or-later |
| 2 | |
| 3 | package main |
| 4 | |
| 5 | import ( |
| 6 | "context" |
| 7 | "fmt" |
| 8 | "log/slog" |
| 9 | "net/netip" |
| 10 | "os" |
| 11 | "os/exec" |
| 12 | "path/filepath" |
| 13 | "strconv" |
| 14 | "strings" |
| 15 | "syscall" |
| 16 | "time" |
| 17 | |
| 18 | "github.com/spf13/cobra" |
| 19 | |
| 20 | "github.com/tenseleyFlow/shithub/internal/git/protocol" |
| 21 | "github.com/tenseleyFlow/shithub/internal/infra/config" |
| 22 | "github.com/tenseleyFlow/shithub/internal/infra/db" |
| 23 | "github.com/tenseleyFlow/shithub/internal/infra/storage" |
| 24 | usersdb "github.com/tenseleyFlow/shithub/internal/users/sqlc" |
| 25 | ) |
| 26 | |
| 27 | // sshAuthkeysCmd implements sshd's AuthorizedKeysCommand contract: |
| 28 | // |
| 29 | // - On a known fingerprint, write a single authorized_keys line on stdout |
| 30 | // with a forced command and restrictive options. |
| 31 | // - On an unknown fingerprint OR any error, write nothing and exit 0. |
| 32 | // sshd uses STDOUT as the auth answer; non-zero exit is a config error, |
| 33 | // not a deny. Failing closed is the right model: better to deny a |
| 34 | // legitimate connection than accidentally authorize the wrong user. |
| 35 | // |
| 36 | // Latency is critical — every SSH connection waits on this. The pool is |
| 37 | // sized small (max 4 conns) to bound startup cost and tail-latency. |
| 38 | var sshAuthkeysCmd = &cobra.Command{ |
| 39 | Use: "ssh-authkeys <fingerprint>", |
| 40 | Short: "AuthorizedKeysCommand handler for sshd", |
| 41 | Args: cobra.ExactArgs(1), |
| 42 | Hidden: true, // not for direct human use |
| 43 | RunE: func(cmd *cobra.Command, args []string) error { |
| 44 | // Fail-closed wrapper: anything below that returns an error or |
| 45 | // panics writes nothing to stdout. The exit code stays 0. |
| 46 | defer func() { |
| 47 | _ = recover() |
| 48 | }() |
| 49 | fp := strings.TrimSpace(args[0]) |
| 50 | if !isWellFormedFingerprint(fp) { |
| 51 | return nil |
| 52 | } |
| 53 | |
| 54 | cfg, err := config.Load(nil) |
| 55 | if err != nil || cfg.DB.URL == "" { |
| 56 | return nil |
| 57 | } |
| 58 | |
| 59 | ctx, cancel := context.WithTimeout(cmd.Context(), 1500*time.Millisecond) |
| 60 | defer cancel() |
| 61 | |
| 62 | pool, err := db.Open(ctx, db.Config{ |
| 63 | URL: cfg.DB.URL, MaxConns: 4, MinConns: 0, |
| 64 | ConnectTimeout: 750 * time.Millisecond, |
| 65 | }) |
| 66 | if err != nil { |
| 67 | return nil |
| 68 | } |
| 69 | defer pool.Close() |
| 70 | |
| 71 | q := usersdb.New() |
| 72 | row, err := q.GetUserSSHKeyByFingerprint(ctx, pool, fp) |
| 73 | if err != nil { |
| 74 | // pgx.ErrNoRows or any other error → silently empty. |
| 75 | return nil |
| 76 | } |
| 77 | |
| 78 | _, _ = fmt.Fprintln(cmd.OutOrStdout(), authorizedKeysLine(row)) |
| 79 | |
| 80 | // Best-effort last-used update. 500ms cap; any error is dropped. |
| 81 | updateCtx, updateCancel := context.WithTimeout(context.Background(), 500*time.Millisecond) |
| 82 | defer updateCancel() |
| 83 | _ = q.TouchSSHKeyLastUsed(updateCtx, pool, usersdb.TouchSSHKeyLastUsedParams{ |
| 84 | ID: row.ID, |
| 85 | LastUsedIp: clientAddrFromEnv(), |
| 86 | }) |
| 87 | return nil |
| 88 | }, |
| 89 | } |
| 90 | |
| 91 | // sshShellCmd is the forced-command target sshd invokes after the |
| 92 | // AuthorizedKeysCommand handshake binds the connection to a user. |
| 93 | // |
| 94 | // Flow on a successful clone/push: |
| 95 | // |
| 96 | // sshd ──► shithubd ssh-shell <user_id> |
| 97 | // ├─ ParseSSHCommand(SSH_ORIGINAL_COMMAND) |
| 98 | // ├─ Resolve user + repo against the DB |
| 99 | // ├─ Inline owner-only authz (S15 will refactor) |
| 100 | // ├─ Build SHITHUB_* env (so post-receive hooks identify the actor) |
| 101 | // ├─ Close the DB pool (syscall.Exec preserves all open FDs) |
| 102 | // └─ syscall.Exec git-{upload,receive}-pack <bare-repo> |
| 103 | // |
| 104 | // On any error: write a friendly line to stderr (the user sees it in |
| 105 | // their git client), log structured, exit non-zero. defer does NOT |
| 106 | // fire on syscall.Exec — every cleanup happens BEFORE the exec call. |
| 107 | var sshShellCmd = &cobra.Command{ |
| 108 | Use: "ssh-shell <user_id>", |
| 109 | Short: "Forced-command target invoked by sshd via AuthorizedKeysCommand", |
| 110 | Args: cobra.ExactArgs(1), |
| 111 | Hidden: true, |
| 112 | RunE: func(cmd *cobra.Command, args []string) error { |
| 113 | userID, err := strconv.ParseInt(args[0], 10, 64) |
| 114 | if err != nil { |
| 115 | _, _ = fmt.Fprintln(cmd.ErrOrStderr(), "shithub: invalid user") |
| 116 | return fmt.Errorf("ssh-shell: bad user_id %q: %w", args[0], err) |
| 117 | } |
| 118 | original := os.Getenv("SSH_ORIGINAL_COMMAND") |
| 119 | remoteIP := protocol.ParseRemoteIP(os.Getenv("SSH_CONNECTION")) |
| 120 | logger := slog.New(slog.NewTextHandler(cmd.ErrOrStderr(), &slog.HandlerOptions{Level: slog.LevelInfo})) |
| 121 | |
| 122 | cfg, err := config.Load(nil) |
| 123 | if err != nil || cfg.DB.URL == "" || cfg.Storage.ReposRoot == "" { |
| 124 | _, _ = fmt.Fprintln(cmd.ErrOrStderr(), "shithub: server misconfigured") |
| 125 | return fmt.Errorf("ssh-shell: cfg: %w", err) |
| 126 | } |
| 127 | root, err := filepath.Abs(cfg.Storage.ReposRoot) |
| 128 | if err != nil { |
| 129 | _, _ = fmt.Fprintln(cmd.ErrOrStderr(), "shithub: server misconfigured") |
| 130 | return fmt.Errorf("ssh-shell: repos_root: %w", err) |
| 131 | } |
| 132 | rfs, err := storage.NewRepoFS(root) |
| 133 | if err != nil { |
| 134 | _, _ = fmt.Fprintln(cmd.ErrOrStderr(), "shithub: server misconfigured") |
| 135 | return fmt.Errorf("ssh-shell: NewRepoFS: %w", err) |
| 136 | } |
| 137 | |
| 138 | ctx, cancel := context.WithTimeout(cmd.Context(), 5*time.Second) |
| 139 | defer cancel() |
| 140 | pool, err := db.Open(ctx, db.Config{ |
| 141 | URL: cfg.DB.URL, MaxConns: 2, MinConns: 0, |
| 142 | ConnectTimeout: 1500 * time.Millisecond, |
| 143 | }) |
| 144 | if err != nil { |
| 145 | _, _ = fmt.Fprintln(cmd.ErrOrStderr(), "shithub: temporary failure (try again)") |
| 146 | return fmt.Errorf("ssh-shell: db open: %w", err) |
| 147 | } |
| 148 | |
| 149 | res, parsed, dispatchErr := protocol.PrepareDispatch(ctx, protocol.SSHDispatchDeps{ |
| 150 | Pool: pool, RepoFS: rfs, |
| 151 | }, protocol.SSHDispatchInput{ |
| 152 | OriginalCommand: original, |
| 153 | UserID: userID, |
| 154 | RemoteIP: remoteIP, |
| 155 | }) |
| 156 | if dispatchErr != nil { |
| 157 | pool.Close() |
| 158 | _, _ = fmt.Fprintln(cmd.ErrOrStderr(), protocol.FriendlyMessageFor(dispatchErr, "")) |
| 159 | logger.WarnContext( |
| 160 | ctx, "ssh-shell: denied", |
| 161 | "user_id", userID, |
| 162 | "original", original, |
| 163 | "remote_ip", remoteIP, |
| 164 | "error", dispatchErr, |
| 165 | ) |
| 166 | return dispatchErr |
| 167 | } |
| 168 | logger.InfoContext( |
| 169 | ctx, "ssh-shell: dispatch", |
| 170 | "user_id", userID, |
| 171 | "op", string(parsed.Service), |
| 172 | "owner", parsed.Owner, |
| 173 | "repo", parsed.Repo, |
| 174 | "remote_ip", remoteIP, |
| 175 | ) |
| 176 | |
| 177 | // CRITICAL: close DB pool before syscall.Exec. defer doesn't |
| 178 | // fire on exec, and the pgx pool's connections would otherwise |
| 179 | // leak into the new process's FD table. |
| 180 | pool.Close() |
| 181 | |
| 182 | bin, err := exec.LookPath(res.Argv0) |
| 183 | if err != nil { |
| 184 | _, _ = fmt.Fprintln(cmd.ErrOrStderr(), "shithub: server misconfigured") |
| 185 | return fmt.Errorf("ssh-shell: lookup %s: %w", res.Argv0, err) |
| 186 | } |
| 187 | if err := sysExec(bin, res.Argv0Args, res.Env); err != nil { |
| 188 | _, _ = fmt.Fprintln(cmd.ErrOrStderr(), "shithub: internal error") |
| 189 | return fmt.Errorf("ssh-shell: exec %s: %w", bin, err) |
| 190 | } |
| 191 | // Unreachable on success — syscall.Exec replaces this process. |
| 192 | return nil |
| 193 | }, |
| 194 | } |
| 195 | |
| 196 | // sysExec is split out so tests can stub it. bin is exec.LookPath of a |
| 197 | // fixed service name (git-{upload,receive}-pack); argv[1] is the |
| 198 | // sanitized bare-repo path from storage.RepoFS. |
| 199 | // |
| 200 | //nolint:gosec // G204: inputs are constrained as documented above. |
| 201 | var sysExec = syscall.Exec |
| 202 | |
| 203 | // authorizedKeysLine builds the single line sshd consumes. The forced |
| 204 | // command runs `shithubd ssh-shell <user_id>`; the option set strips |
| 205 | // every interactive affordance. |
| 206 | // |
| 207 | // The command uses the BARE binary name (`shithubd`), not the absolute |
| 208 | // path of os.Args[0]. Reason: when the git user's login shell is |
| 209 | // /usr/bin/git-shell (kept as a defense-in-depth wall — git-shell |
| 210 | // only allows git's own commands plus entries in ~/git-shell-commands/), |
| 211 | // git-shell rejects any first token that contains a slash. With a |
| 212 | // bare token, git-shell looks for ~git/git-shell-commands/shithubd |
| 213 | // (must be a symlink to /usr/local/bin/shithubd installed by ansible) |
| 214 | // and execs it. |
| 215 | // |
| 216 | // PATH on the AKC's child process inherits sshd's env, which on |
| 217 | // Debian includes /usr/local/bin — so a fallback `shithubd` lookup |
| 218 | // would also find the binary even without git-shell-commands. The |
| 219 | // git-shell-commands symlink is the authoritative path. |
| 220 | func authorizedKeysLine(row usersdb.UserSshKey) string { |
| 221 | binary := filepath.Base(os.Args[0]) |
| 222 | // user_id is a digit string so it can never contain shell metacharacters. |
| 223 | cmd := fmt.Sprintf(`%s ssh-shell %d`, binary, row.UserID) |
| 224 | options := strings.Join([]string{ |
| 225 | fmt.Sprintf(`command="%s"`, cmd), |
| 226 | "no-port-forwarding", |
| 227 | "no-X11-forwarding", |
| 228 | "no-agent-forwarding", |
| 229 | "no-pty", |
| 230 | }, ",") |
| 231 | return options + " " + row.PublicKey |
| 232 | } |
| 233 | |
| 234 | // clientAddrFromEnv extracts the connecting client's address from |
| 235 | // $SSH_CONNECTION (sshd sets it to "<client> <cport> <server> <sport>"). |
| 236 | // Returns nil when unavailable, which sqlc encodes as a SQL NULL. |
| 237 | func clientAddrFromEnv() *netip.Addr { |
| 238 | conn := os.Getenv("SSH_CONNECTION") |
| 239 | if conn == "" { |
| 240 | return nil |
| 241 | } |
| 242 | parts := strings.Fields(conn) |
| 243 | if len(parts) < 1 { |
| 244 | return nil |
| 245 | } |
| 246 | addr, err := netip.ParseAddr(parts[0]) |
| 247 | if err != nil { |
| 248 | return nil |
| 249 | } |
| 250 | return &addr |
| 251 | } |
| 252 | |
| 253 | // isWellFormedFingerprint accepts only the canonical SHA256:<b64> shape |
| 254 | // our codebase emits. Defense against an attacker passing crafted strings |
| 255 | // to influence the SQL plan. |
| 256 | func isWellFormedFingerprint(s string) bool { |
| 257 | if !strings.HasPrefix(s, "SHA256:") { |
| 258 | return false |
| 259 | } |
| 260 | rest := s[len("SHA256:"):] |
| 261 | if len(rest) < 30 || len(rest) > 80 { |
| 262 | return false |
| 263 | } |
| 264 | for _, r := range rest { |
| 265 | switch { |
| 266 | case r >= 'A' && r <= 'Z', |
| 267 | r >= 'a' && r <= 'z', |
| 268 | r >= '0' && r <= '9', |
| 269 | r == '+', r == '/', r == '=': |
| 270 | default: |
| 271 | return false |
| 272 | } |
| 273 | } |
| 274 | return true |
| 275 | } |
| 276 | |
| 277 | func init() { |
| 278 | rootCmd.AddCommand(sshAuthkeysCmd) |
| 279 | rootCmd.AddCommand(sshShellCmd) |
| 280 | } |
| 281 |