container: optionally map uid/gid 0 as init

Unfortunately required to work around flawed APIs like binfmt_misc.

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

container/container_test.go

@@ -409,8 +409,11 @@ var containerTestCases = []struct {
 func TestContainer(t *testing.T) {
 	t.Parallel()
 
+	var suffix string
+runTests:
 	for i, tc := range containerTestCases {
-		t.Run(tc.name, func(t *testing.T) {
+		_suffix := suffix
+		t.Run(tc.name+_suffix, func(t *testing.T) {
 			t.Parallel()
 
 			wantOps, wantOpsCtx := tc.ops(t)
@@ -434,6 +437,8 @@ func TestContainer(t *testing.T) {
 			c.SeccompDisable = !tc.filter
 			c.RetainSession = tc.session
 			c.HostNet = tc.net
+			c.InitAsRoot = _suffix != ""
+			c.Env = append(c.Env, "HAKUREI_TEST_SUFFIX="+_suffix)
 			if info.CanDegrade {
 				if _, err := landlock.GetABI(); err != nil {
 					if !errors.Is(err, syscall.ENOSYS) {
@@ -443,6 +448,9 @@ func TestContainer(t *testing.T) {
 					t.Log("Landlock LSM is unavailable, enabling HostAbstract")
 				}
 			}
+			if c.InitAsRoot {
+				c.SeccompPresets &= ^std.PresetDenyNS
+			}
 
 			c.
 				Readonly(check.MustAbs(pathReadonly), 0755).
@@ -511,6 +519,11 @@ func TestContainer(t *testing.T) {
 			}
 		})
 	}
+
+	if suffix == "" {
+		suffix = " as root"
+		goto runTests
+	}
 }
 
 func ent(root, target, vfsOptstr, fsType, source, fsOptstr string) *vfs.MountInfoEntry {
@@ -589,9 +602,9 @@ func testContainerCancel(
 }
 
 func TestForward(t *testing.T) {
-	testContainerCancel(t, func(c *container.Container) {
-		c.ForwardCancel = true
-	}, func(ps *os.ProcessState, waitErr error) {
+	t.Parallel()
+
+	f := func(ps *os.ProcessState, waitErr error) {
 		var exitError *exec.ExitError
 		if !errors.As(waitErr, &exitError) {
 			if m, ok := container.InternalMessageFromError(waitErr); ok {
@@ -602,11 +615,26 @@ func TestForward(t *testing.T) {
 		if code := exitError.ExitCode(); code != blockExitCodeInterrupt {
 			t.Errorf("ExitCode: %d, want %d", code, blockExitCodeInterrupt)
 		}
+	}
+	t.Run("direct", func(t *testing.T) {
+		t.Parallel()
+		testContainerCancel(t, func(c *container.Container) {
+			c.ForwardCancel = true
+		}, f)
+	})
+	t.Run("as root", func(t *testing.T) {
+		testContainerCancel(t, func(c *container.Container) {
+			c.ForwardCancel = true
+			c.InitAsRoot = true
+			c.Proc(fhs.AbsProc)
+		}, f)
 	})
 }
 
 func TestCancel(t *testing.T) {
-	testContainerCancel(t, nil, func(ps *os.ProcessState, waitErr error) {
+	t.Parallel()
+
+	f := func(ps *os.ProcessState, waitErr error) {
 		wantErr := context.Canceled
 		if !reflect.DeepEqual(waitErr, wantErr) {
 			if m, ok := container.InternalMessageFromError(waitErr); ok {
@@ -619,6 +647,16 @@ func TestCancel(t *testing.T) {
 		} else if code := ps.ExitCode(); code != 0 {
 			t.Errorf("ExitCode: %d, want %d", code, 0)
 		}
+	}
+	t.Run("direct", func(t *testing.T) {
+		t.Parallel()
+		testContainerCancel(t, nil, f)
+	})
+	t.Run("as root", func(t *testing.T) {
+		testContainerCancel(t, func(c *container.Container) {
+			c.InitAsRoot = true
+			c.Proc(fhs.AbsProc)
+		}, f)
 	})
 }
 
@@ -655,6 +693,8 @@ func init() {
 		})
 
 		c.Command("container", command.UsageInternal, func(args []string) error {
+			asRoot := os.Getenv("HAKUREI_TEST_SUFFIX") == " as root"
+
 			if len(args) != 1 {
 				return syscall.EINVAL
 			}
@@ -672,11 +712,19 @@ func init() {
 				return fmt.Errorf("gid: %d, want %d", gid, tc.gid)
 			}
 
+			// no attack surface increase during as root due to no_new_privs
+			var wantBounding uintptr = 1
+			asRootNot := " not"
+			if !asRoot {
+				wantBounding = 0
+				asRootNot = ""
+			}
+
 			const (
 				PR_CAP_AMBIENT        = 0x2f
 				PR_CAP_AMBIENT_IS_SET = 0x1
 			)
-			for i := range container.LastCap(nil) {
+			for i := range container.LastCap(nil) + 1 {
 				r, _, errno := syscall.Syscall(
 					syscall.SYS_PRCTL,
 					PR_CAP_AMBIENT,
@@ -687,7 +735,7 @@ func init() {
 					return os.NewSyscallError("prctl", errno)
 				}
 				if r != 0 {
-					return fmt.Errorf("capability %d is set", i)
+					return fmt.Errorf("capability %d in ambient set", i)
 				}
 
 				r, _, errno = syscall.Syscall(
@@ -699,8 +747,8 @@ func init() {
 				if errno != 0 {
 					return os.NewSyscallError("prctl", errno)
 				}
-				if r != 0 {
-					return fmt.Errorf("capability %d in set", i)
+				if r != wantBounding {
+					return fmt.Errorf("capability %d%s in bounding set", i, asRootNot)
 				}
 			}