container: set FD_CLOEXEC on all open files

While fd created from this side always has the FD_CLOEXEC flag, the same is not true for files left open by the parent. This change prevents those files from leaking into the container.

Signed-off-by: Ophestra <cat@gensokyo.uk>

container/container.go

@@ -11,6 +11,7 @@ import (
 	"os/exec"
 	"runtime"
 	"strconv"
+	"sync"
 	. "syscall"
 	"time"
 
@@ -143,11 +144,18 @@ func (e *StartError) Error() string {
 // Message returns a user-facing error message.
 func (e *StartError) Message() string {
 	if e.Passthrough {
+		var (
+			numError *strconv.NumError
+		)
+
 		switch {
 		case errors.As(e.Err, new(*os.PathError)),
 			errors.As(e.Err, new(*os.SyscallError)):
 			return "cannot " + e.Err.Error()
 
+		case errors.As(e.Err, &numError) && numError != nil:
+			return "cannot parse " + strconv.Quote(numError.Num) + ": " + numError.Err.Error()
+
 		default:
 			return e.Err.Error()
 		}
@@ -158,6 +166,39 @@ func (e *StartError) Message() string {
 	return "cannot " + e.Error()
 }
 
+// for ensureCloseOnExec
+var (
+	closeOnExecOnce sync.Once
+	closeOnExecErr  error
+)
+
+// ensureCloseOnExec ensures all currently open file descriptors have the syscall.FD_CLOEXEC flag set.
+// This is only ran once as it is intended to handle files left open by the parent, and any file opened
+// on this side should already have syscall.FD_CLOEXEC set.
+func ensureCloseOnExec() error {
+	closeOnExecOnce.Do(func() {
+		const fdPrefixPath = "/proc/self/fd/"
+
+		var entries []os.DirEntry
+		if entries, closeOnExecErr = os.ReadDir(fdPrefixPath); closeOnExecErr != nil {
+			return
+		}
+
+		var fd int
+		for _, ent := range entries {
+			if fd, closeOnExecErr = strconv.Atoi(ent.Name()); closeOnExecErr != nil {
+				break // not reached
+			}
+			CloseOnExec(fd)
+		}
+	})
+
+	if closeOnExecErr == nil {
+		return nil
+	}
+	return &StartError{Fatal: true, Step: "set FD_CLOEXEC on all open files", Err: closeOnExecErr, Passthrough: true}
+}
+
 // Start starts the container init. The init process blocks until Serve is called.
 func (p *Container) Start() error {
 	if p == nil || p.cmd == nil ||
@@ -168,6 +209,10 @@ func (p *Container) Start() error {
 		return errors.New("container: already started")
 	}
 
+	if err := ensureCloseOnExec(); err != nil {
+		return err
+	}
+
 	// map to overflow id to work around ownership checks
 	if p.Uid < 1 {
 		p.Uid = OverflowUid(p.msg)

container/container_test.go

@@ -44,8 +44,7 @@ func TestStartError(t *testing.T) {
 			Fatal: true,
 			Step:  "set up params stream",
 			Err:   container.ErrReceiveEnv,
-		},
+		}, "set up params stream: environment variable not set",
-			"set up params stream: environment variable not set",
 			container.ErrReceiveEnv, syscall.EBADF,
 			"cannot set up params stream: environment variable not set"},
 
@@ -53,8 +52,7 @@ func TestStartError(t *testing.T) {
 			Fatal: true,
 			Step:  "set up params stream",
 			Err:   &os.SyscallError{Syscall: "pipe2", Err: syscall.EBADF},
-		},
+		}, "set up params stream pipe2: bad file descriptor",
-			"set up params stream pipe2: bad file descriptor",
 			syscall.EBADF, os.ErrInvalid,
 			"cannot set up params stream pipe2: bad file descriptor"},
 
@@ -62,16 +60,14 @@ func TestStartError(t *testing.T) {
 			Fatal: true,
 			Step:  "prctl(PR_SET_NO_NEW_PRIVS)",
 			Err:   syscall.EPERM,
-		},
+		}, "prctl(PR_SET_NO_NEW_PRIVS): operation not permitted",
-			"prctl(PR_SET_NO_NEW_PRIVS): operation not permitted",
 			syscall.EPERM, syscall.EACCES,
 			"cannot prctl(PR_SET_NO_NEW_PRIVS): operation not permitted"},
 
 		{"landlock abi", &container.StartError{
 			Step: "get landlock ABI",
 			Err:  syscall.ENOSYS,
-		},
+		}, "get landlock ABI: function not implemented",
-			"get landlock ABI: function not implemented",
 			syscall.ENOSYS, syscall.ENOEXEC,
 			"cannot get landlock ABI: function not implemented"},
 
@@ -79,8 +75,7 @@ func TestStartError(t *testing.T) {
 			Step:   "kernel version too old for LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET",
 			Err:    syscall.ENOSYS,
 			Origin: true,
-		},
+		}, "kernel version too old for LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET",
-			"kernel version too old for LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET",
 			syscall.ENOSYS, syscall.ENOSPC,
 			"kernel version too old for LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET"},
 
@@ -88,8 +83,7 @@ func TestStartError(t *testing.T) {
 			Fatal: true,
 			Step:  "create landlock ruleset",
 			Err:   syscall.EBADFD,
-		},
+		}, "create landlock ruleset: file descriptor in bad state",
-			"create landlock ruleset: file descriptor in bad state",
 			syscall.EBADFD, syscall.EBADF,
 			"cannot create landlock ruleset: file descriptor in bad state"},
 
@@ -97,8 +91,7 @@ func TestStartError(t *testing.T) {
 			Fatal: true,
 			Step:  "enforce landlock ruleset",
 			Err:   syscall.ENOTRECOVERABLE,
-		},
+		}, "enforce landlock ruleset: state not recoverable",
-			"enforce landlock ruleset: state not recoverable",
 			syscall.ENOTRECOVERABLE, syscall.ETIMEDOUT,
 			"cannot enforce landlock ruleset: state not recoverable"},
 
@@ -109,8 +102,7 @@ func TestStartError(t *testing.T) {
 				Path: "/proc/nonexistent",
 				Err:  syscall.ENOENT,
 			}, Passthrough: true,
-		},
+		}, "fork/exec /proc/nonexistent: no such file or directory",
-			"fork/exec /proc/nonexistent: no such file or directory",
 			syscall.ENOENT, syscall.ENOSYS,
 			"cannot fork/exec /proc/nonexistent: no such file or directory"},
 
@@ -120,11 +112,19 @@ func TestStartError(t *testing.T) {
 				Syscall: "open",
 				Err:     syscall.ENOSYS,
 			}, Passthrough: true,
-		},
+		}, "open: function not implemented",
-			"open: function not implemented",
 			syscall.ENOSYS, syscall.ENOENT,
 			"cannot open: function not implemented"},
 
+		{"start FD_CLOEXEC", &container.StartError{
+			Fatal:       true,
+			Step:        "set FD_CLOEXEC on all open files",
+			Err:         func() error { _, err := strconv.Atoi("invalid"); return err }(),
+			Passthrough: true,
+		}, `strconv.Atoi: parsing "invalid": invalid syntax`,
+			strconv.ErrSyntax, os.ErrInvalid,
+			`cannot parse "invalid": invalid syntax`},
+
 		{"start other", &container.StartError{
 			Step: "start container init",
 			Err: &net.OpError{
@@ -132,8 +132,7 @@ func TestStartError(t *testing.T) {
 				Net: "unix",
 				Err: syscall.ECONNREFUSED,
 			}, Passthrough: true,
-		},
+		}, "dial unix: connection refused",
-			"dial unix: connection refused",
 			syscall.ECONNREFUSED, syscall.ECONNABORTED,
 			"dial unix: connection refused"},
 	}

test/sandbox/test.py

@@ -46,7 +46,7 @@ swaymsg("exec hakurei run cat")
 check_filter(0, "pdlike", "cat")
 
 # Check fd leak:
-swaymsg("exec hakurei -v run sleep infinity")
+swaymsg("exec exec 127</proc/cmdline && hakurei -v run sleep infinity")
 pd_identity0_sleep_pid = int(machine.wait_until_succeeds("pgrep -U 10000 -x sleep", timeout=60))
 print(machine.succeed(f"hakurei-test fd {pd_identity0_sleep_pid}"))
 machine.succeed(f"kill -INT {pd_identity0_sleep_pid}")