container: optionally isolate host abstract UNIX domain sockets via landlock

17 files changed, 375 insertions, 9 deletions

diff --git a/cmd/hpkg/app.go b/cmd/hpkg/app.go
index 67d12496..713a4fc7 100644
--- a/cmd/hpkg/app.go
+++ b/cmd/hpkg/app.go
@@ -28,6 +28,8 @@ type appInfo struct {
 	// passed through to [hst.Config]
 	Net bool `json:"net,omitempty"`
 	// passed through to [hst.Config]
+	Abstract bool `json:"abstract,omitempty"`
+	// passed through to [hst.Config]
 	Device bool `json:"dev,omitempty"`
 	// passed through to [hst.Config]
 	Tty bool `json:"tty,omitempty"`
@@ -87,6 +89,7 @@ func (app *appInfo) toHst(pathSet *appPathSet, pathname *container.Absolute, arg
 			Devel:      app.Devel,
 			Userns:     app.Userns,
 			Net:        app.Net,
+			Abstract:   app.Abstract,
 			Device:     app.Device,
 			Tty:        app.Tty || flagDropShell,
 			MapRealUID: app.MapRealUID,
diff --git a/container/container.go b/container/container.go
index 52691cea..fa5a34e5 100644
--- a/container/container.go
+++ b/container/container.go
@@ -92,6 +92,8 @@ type (
 		RetainSession bool
 		// Do not [syscall.CLONE_NEWNET].
 		HostNet bool
+		// Do not [LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET].
+		HostAbstract bool
 		// Retain CAP_SYS_ADMIN.
 		Privileged bool
 	}
@@ -185,6 +187,51 @@ func (p *Container) Start() error {
 					"prctl(PR_SET_NO_NEW_PRIVS):")
 			}
 
+			// landlock: depends on per-thread state but acts on a process group
+			{
+				rulesetAttr := &RulesetAttr{Scoped: LANDLOCK_SCOPE_SIGNAL}
+				if !p.HostAbstract {
+					rulesetAttr.Scoped |= LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET
+				}
+
+				if abi, err := LandlockGetABI(); err != nil {
+					if p.HostAbstract {
+						// landlock can be skipped here as it restricts access to resources
+						// already covered by namespaces (pid)
+						goto landlockOut
+					}
+					return wrapErrSuffix(err,
+						"landlock does not appear to be enabled:")
+				} else if abi < 6 {
+					if p.HostAbstract {
+						// see above comment
+						goto landlockOut
+					}
+					return msg.WrapErr(ENOSYS,
+						"kernel version too old for LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET")
+				} else {
+					msg.Verbosef("landlock abi version %d", abi)
+				}
+
+				if rulesetFd, err := rulesetAttr.Create(0); err != nil {
+					return wrapErrSuffix(err,
+						"cannot create landlock ruleset:")
+				} else {
+					msg.Verbosef("enforcing landlock ruleset %s", rulesetAttr)
+					if err = LandlockRestrictSelf(rulesetFd, 0); err != nil {
+						_ = Close(rulesetFd)
+						return wrapErrSuffix(err,
+							"cannot enforce landlock ruleset:")
+					}
+					if err = Close(rulesetFd); err != nil {
+						msg.Verbosef("cannot close landlock ruleset: %v", err)
+						// not fatal
+					}
+				}
+
+			landlockOut:
+			}
+
 			msg.Verbose("starting container init")
 			if err := p.cmd.Start(); err != nil {
 				return msg.WrapErr(err, err.Error())
diff --git a/container/landlock.go b/container/landlock.go
new file mode 100644
index 00000000..7f7461e2
--- /dev/null
+++ b/container/landlock.go
@@ -0,0 +1,239 @@
+package container
+
+import (
+	"strings"
+	"syscall"
+	"unsafe"
+
+	"hakurei.app/container/seccomp"
+)
+
+// include/uapi/linux/landlock.h
+
+const (
+	LANDLOCK_CREATE_RULESET_VERSION = 1 << iota
+)
+
+type LandlockAccessFS uintptr
+
+const (
+	LANDLOCK_ACCESS_FS_EXECUTE LandlockAccessFS = 1 << iota
+	LANDLOCK_ACCESS_FS_WRITE_FILE
+	LANDLOCK_ACCESS_FS_READ_FILE
+	LANDLOCK_ACCESS_FS_READ_DIR
+	LANDLOCK_ACCESS_FS_REMOVE_DIR
+	LANDLOCK_ACCESS_FS_REMOVE_FILE
+	LANDLOCK_ACCESS_FS_MAKE_CHAR
+	LANDLOCK_ACCESS_FS_MAKE_DIR
+	LANDLOCK_ACCESS_FS_MAKE_REG
+	LANDLOCK_ACCESS_FS_MAKE_SOCK
+	LANDLOCK_ACCESS_FS_MAKE_FIFO
+	LANDLOCK_ACCESS_FS_MAKE_BLOCK
+	LANDLOCK_ACCESS_FS_MAKE_SYM
+	LANDLOCK_ACCESS_FS_REFER
+	LANDLOCK_ACCESS_FS_TRUNCATE
+	LANDLOCK_ACCESS_FS_IOCTL_DEV
+
+	_LANDLOCK_ACCESS_FS_DELIM
+)
+
+func (f LandlockAccessFS) String() string {
+	switch f {
+	case LANDLOCK_ACCESS_FS_EXECUTE:
+		return "execute"
+
+	case LANDLOCK_ACCESS_FS_WRITE_FILE:
+		return "write_file"
+
+	case LANDLOCK_ACCESS_FS_READ_FILE:
+		return "read_file"
+
+	case LANDLOCK_ACCESS_FS_READ_DIR:
+		return "read_dir"
+
+	case LANDLOCK_ACCESS_FS_REMOVE_DIR:
+		return "remove_dir"
+
+	case LANDLOCK_ACCESS_FS_REMOVE_FILE:
+		return "remove_file"
+
+	case LANDLOCK_ACCESS_FS_MAKE_CHAR:
+		return "make_char"
+
+	case LANDLOCK_ACCESS_FS_MAKE_DIR:
+		return "make_dir"
+
+	case LANDLOCK_ACCESS_FS_MAKE_REG:
+		return "make_reg"
+
+	case LANDLOCK_ACCESS_FS_MAKE_SOCK:
+		return "make_sock"
+
+	case LANDLOCK_ACCESS_FS_MAKE_FIFO:
+		return "make_fifo"
+
+	case LANDLOCK_ACCESS_FS_MAKE_BLOCK:
+		return "make_block"
+
+	case LANDLOCK_ACCESS_FS_MAKE_SYM:
+		return "make_sym"
+
+	case LANDLOCK_ACCESS_FS_REFER:
+		return "fs_refer"
+
+	case LANDLOCK_ACCESS_FS_TRUNCATE:
+		return "fs_truncate"
+
+	case LANDLOCK_ACCESS_FS_IOCTL_DEV:
+		return "fs_ioctl_dev"
+
+	default:
+		var c []LandlockAccessFS
+		for i := LandlockAccessFS(1); i < _LANDLOCK_ACCESS_FS_DELIM; i <<= 1 {
+			if f&i != 0 {
+				c = append(c, i)
+			}
+		}
+		if len(c) == 0 {
+			return "NULL"
+		}
+		s := make([]string, len(c))
+		for i, v := range c {
+			s[i] = v.String()
+		}
+		return strings.Join(s, " ")
+	}
+}
+
+type LandlockAccessNet uintptr
+
+const (
+	LANDLOCK_ACCESS_NET_BIND_TCP LandlockAccessNet = 1 << iota
+	LANDLOCK_ACCESS_NET_CONNECT_TCP
+
+	_LANDLOCK_ACCESS_NET_DELIM
+)
+
+func (f LandlockAccessNet) String() string {
+	switch f {
+	case LANDLOCK_ACCESS_NET_BIND_TCP:
+		return "bind_tcp"
+
+	case LANDLOCK_ACCESS_NET_CONNECT_TCP:
+		return "connect_tcp"
+
+	default:
+		var c []LandlockAccessNet
+		for i := LandlockAccessNet(1); i < _LANDLOCK_ACCESS_NET_DELIM; i <<= 1 {
+			if f&i != 0 {
+				c = append(c, i)
+			}
+		}
+		if len(c) == 0 {
+			return "NULL"
+		}
+		s := make([]string, len(c))
+		for i, v := range c {
+			s[i] = v.String()
+		}
+		return strings.Join(s, " ")
+	}
+}
+
+type LandlockScope uintptr
+
+const (
+	LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET LandlockScope = 1 << iota
+	LANDLOCK_SCOPE_SIGNAL
+
+	_LANDLOCK_SCOPE_DELIM
+)
+
+func (f LandlockScope) String() string {
+	switch f {
+	case LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET:
+		return "abstract_unix_socket"
+
+	case LANDLOCK_SCOPE_SIGNAL:
+		return "signal"
+
+	default:
+		var c []LandlockScope
+		for i := LandlockScope(1); i < _LANDLOCK_SCOPE_DELIM; i <<= 1 {
+			if f&i != 0 {
+				c = append(c, i)
+			}
+		}
+		if len(c) == 0 {
+			return "NULL"
+		}
+		s := make([]string, len(c))
+		for i, v := range c {
+			s[i] = v.String()
+		}
+		return strings.Join(s, " ")
+	}
+}
+
+type RulesetAttr struct {
+	// Bitmask of handled filesystem actions.
+	HandledAccessFS LandlockAccessFS
+	// Bitmask of handled network actions.
+	HandledAccessNet LandlockAccessNet
+	// Bitmask of scopes restricting a Landlock domain from accessing outside resources (e.g. IPCs).
+	Scoped LandlockScope
+}
+
+func (rulesetAttr *RulesetAttr) String() string {
+	if rulesetAttr == nil {
+		return "NULL"
+	}
+	elems := make([]string, 0, 3)
+	if rulesetAttr.HandledAccessFS > 0 {
+		elems = append(elems, "fs: "+rulesetAttr.HandledAccessFS.String())
+	}
+	if rulesetAttr.HandledAccessNet > 0 {
+		elems = append(elems, "net: "+rulesetAttr.HandledAccessNet.String())
+	}
+	if rulesetAttr.Scoped > 0 {
+		elems = append(elems, "scoped: "+rulesetAttr.Scoped.String())
+	}
+	if len(elems) == 0 {
+		return "0"
+	}
+	return strings.Join(elems, ", ")
+}
+
+func (rulesetAttr *RulesetAttr) Create(flags uintptr) (fd int, err error) {
+	var pointer, size uintptr
+	// NULL needed for abi version
+	if rulesetAttr != nil {
+		pointer = uintptr(unsafe.Pointer(rulesetAttr))
+		size = unsafe.Sizeof(*rulesetAttr)
+	}
+
+	rulesetFd, _, errno := syscall.Syscall(seccomp.SYS_LANDLOCK_CREATE_RULESET, pointer, size, flags)
+	fd = int(rulesetFd)
+	err = errno
+
+	if fd < 0 {
+		return
+	}
+
+	if rulesetAttr != nil { // not a fd otherwise
+		syscall.CloseOnExec(fd)
+	}
+	return fd, nil
+}
+
+func LandlockGetABI() (int, error) {
+	return (*RulesetAttr)(nil).Create(LANDLOCK_CREATE_RULESET_VERSION)
+}
+
+func LandlockRestrictSelf(rulesetFd int, flags uintptr) error {
+	r, _, errno := syscall.Syscall(seccomp.SYS_LANDLOCK_RESTRICT_SELF, uintptr(rulesetFd), flags, 0)
+	if r != 0 {
+		return errno
+	}
+	return nil
+}
diff --git a/container/landlock_test.go b/container/landlock_test.go
new file mode 100644
index 00000000..d204bc31
--- /dev/null
+++ b/container/landlock_test.go
@@ -0,0 +1,61 @@
+package container_test
+
+import (
+	"testing"
+	"unsafe"
+
+	"hakurei.app/container"
+)
+
+func TestLandlockString(t *testing.T) {
+	testCases := []struct {
+		name        string
+		rulesetAttr *container.RulesetAttr
+		want        string
+	}{
+		{"nil", nil, "NULL"},
+		{"zero", new(container.RulesetAttr), "0"},
+		{"some", &container.RulesetAttr{Scoped: container.LANDLOCK_SCOPE_SIGNAL}, "scoped: signal"},
+		{"set", &container.RulesetAttr{
+			HandledAccessFS:  container.LANDLOCK_ACCESS_FS_MAKE_SYM | container.LANDLOCK_ACCESS_FS_IOCTL_DEV | container.LANDLOCK_ACCESS_FS_WRITE_FILE,
+			HandledAccessNet: container.LANDLOCK_ACCESS_NET_BIND_TCP,
+			Scoped:           container.LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET | container.LANDLOCK_SCOPE_SIGNAL,
+		}, "fs: write_file make_sym fs_ioctl_dev, net: bind_tcp, scoped: abstract_unix_socket signal"},
+		{"all", &container.RulesetAttr{
+			HandledAccessFS: container.LANDLOCK_ACCESS_FS_EXECUTE |
+				container.LANDLOCK_ACCESS_FS_WRITE_FILE |
+				container.LANDLOCK_ACCESS_FS_READ_FILE |
+				container.LANDLOCK_ACCESS_FS_READ_DIR |
+				container.LANDLOCK_ACCESS_FS_REMOVE_DIR |
+				container.LANDLOCK_ACCESS_FS_REMOVE_FILE |
+				container.LANDLOCK_ACCESS_FS_MAKE_CHAR |
+				container.LANDLOCK_ACCESS_FS_MAKE_DIR |
+				container.LANDLOCK_ACCESS_FS_MAKE_REG |
+				container.LANDLOCK_ACCESS_FS_MAKE_SOCK |
+				container.LANDLOCK_ACCESS_FS_MAKE_FIFO |
+				container.LANDLOCK_ACCESS_FS_MAKE_BLOCK |
+				container.LANDLOCK_ACCESS_FS_MAKE_SYM |
+				container.LANDLOCK_ACCESS_FS_REFER |
+				container.LANDLOCK_ACCESS_FS_TRUNCATE |
+				container.LANDLOCK_ACCESS_FS_IOCTL_DEV,
+			HandledAccessNet: container.LANDLOCK_ACCESS_NET_BIND_TCP |
+				container.LANDLOCK_ACCESS_NET_CONNECT_TCP,
+			Scoped: container.LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET |
+				container.LANDLOCK_SCOPE_SIGNAL,
+		}, "fs: execute write_file read_file read_dir remove_dir remove_file make_char make_dir make_reg make_sock make_fifo make_block make_sym fs_refer fs_truncate fs_ioctl_dev, net: bind_tcp connect_tcp, scoped: abstract_unix_socket signal"},
+	}
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := tc.rulesetAttr.String(); got != tc.want {
+				t.Errorf("String: %s, want %s", got, tc.want)
+			}
+		})
+	}
+}
+
+func TestLandlockAttrSize(t *testing.T) {
+	want := 24
+	if got := unsafe.Sizeof(container.RulesetAttr{}); got != uintptr(want) {
+		t.Errorf("Sizeof: %d, want %d", got, want)
+	}
+}
diff --git a/hst/config.go b/hst/config.go
index 34100e4f..d6f07be0 100644
--- a/hst/config.go
+++ b/hst/config.go
@@ -79,6 +79,8 @@ type (
 		Userns bool `json:"userns,omitempty"`
 		// share host net namespace
 		Net bool `json:"net,omitempty"`
+		// share abstract unix socket scope
+		Abstract bool `json:"abstract,omitempty"`
 		// allow dangerous terminal I/O
 		Tty bool `json:"tty,omitempty"`
 		// allow multiarch
diff --git a/internal/app/app_pd_linux_test.go b/internal/app/app_pd_linux_test.go
index 109e8816..96182d31 100644
--- a/internal/app/app_pd_linux_test.go
+++ b/internal/app/app_pd_linux_test.go
@@ -62,6 +62,7 @@ var testCasesPd = []sealTestCase{
 				Remount(m("/"), syscall.MS_RDONLY),
 			SeccompPresets: seccomp.PresetExt | seccomp.PresetDenyDevel,
 			HostNet:        true,
+			HostAbstract:   true,
 			RetainSession:  true,
 			ForwardCancel:  true,
 		},
@@ -203,6 +204,7 @@ var testCasesPd = []sealTestCase{
 				Remount(m("/"), syscall.MS_RDONLY),
 			SeccompPresets: seccomp.PresetExt | seccomp.PresetDenyDevel,
 			HostNet:        true,
+			HostAbstract:   true,
 			RetainSession:  true,
 			ForwardCancel:  true,
 		},
diff --git a/internal/app/container_linux.go b/internal/app/container_linux.go
index 3f1259fd..7f088192 100644
--- a/internal/app/container_linux.go
+++ b/internal/app/container_linux.go
@@ -33,6 +33,7 @@ func newContainer(s *hst.ContainerConfig, os sys.State, prefix string, uid, gid
 		SeccompPresets: s.SeccompPresets,
 		RetainSession:  s.Tty,
 		HostNet:        s.Net,
+		HostAbstract:   s.Abstract,
 
 		// the container is canceled when shim is requested to exit or receives an interrupt or termination signal;
 		// this behaviour is implemented in the shim
diff --git a/internal/app/seal_linux.go b/internal/app/seal_linux.go
index c7293720..0ae121fb 100644
--- a/internal/app/seal_linux.go
+++ b/internal/app/seal_linux.go
@@ -238,10 +238,11 @@ func (seal *outcome) finalise(ctx context.Context, sys sys.State, config *hst.Co
 		}
 
 		conf := &hst.ContainerConfig{
-			Userns:  true,
-			Net:     true,
-			Tty:     true,
-			AutoEtc: true,
+			Userns:   true,
+			Net:      true,
+			Abstract: true,
+			Tty:      true,
+			AutoEtc:  true,
 
 			AutoRoot:  container.AbsFHSRoot,
 			RootFlags: container.BindWritable,
diff --git a/nixos.nix b/nixos.nix
index 4f083bcb..eb8df5f8 100644
--- a/nixos.nix
+++ b/nixos.nix
@@ -132,6 +132,7 @@ in
                         devel
                         userns
                         net
+                        abstract
                         device
                         tty
                         multiarch
diff --git a/options.nix b/options.nix
index a51609c9..ec2533c0 100644
--- a/options.nix
+++ b/options.nix
@@ -182,6 +182,7 @@ in
               net = mkEnableOption "network access" // {
                 default = true;
               };
+              abstract = mkEnableOption "abstract unix domain socket access";
 
               nix = mkEnableOption "nix daemon access";
               mapRealUid = mkEnableOption "mapping to priv-user uid";
diff --git a/system/dbus/proc.go b/system/dbus/proc.go
index 0e5f1cef..6607ef19 100644
--- a/system/dbus/proc.go
+++ b/system/dbus/proc.go
@@ -64,6 +64,10 @@ func (p *Proxy) Start() error {
 			argF, func(z *container.Container) {
 				z.SeccompFlags |= seccomp.AllowMultiarch
 				z.SeccompPresets |= seccomp.PresetStrict
+
+				// xdg-dbus-proxy fails with scoped abstract unix sockets despite pathname socket being available
+				z.HostAbstract = true
+
 				z.Hostname = "hakurei-dbus"
 				if p.output != nil {
 					z.Stdout, z.Stderr = p.output, p.output
diff --git a/test/sandbox/case/device.nix b/test/sandbox/case/device.nix
index e9cf724b..c1a2f5f5 100644
--- a/test/sandbox/case/device.nix
+++ b/test/sandbox/case/device.nix
@@ -243,7 +243,7 @@ in
     seccomp = true;
 
     try_socket = "/tmp/.X11-unix/X0";
-    socket_abstract = true;
+    socket_abstract = false;
     socket_pathname = true;
   };
 }
diff --git a/test/sandbox/case/mapuid.nix b/test/sandbox/case/mapuid.nix
index 18d82e77..3ef5f247 100644
--- a/test/sandbox/case/mapuid.nix
+++ b/test/sandbox/case/mapuid.nix
@@ -269,7 +269,7 @@ in
     seccomp = true;
 
     try_socket = "/tmp/.X11-unix/X0";
-    socket_abstract = true;
+    socket_abstract = false;
     socket_pathname = false;
   };
 }
diff --git a/test/sandbox/case/pd.nix b/test/sandbox/case/pd.nix
index 79400f53..49230bcf 100644
--- a/test/sandbox/case/pd.nix
+++ b/test/sandbox/case/pd.nix
@@ -194,5 +194,9 @@
     ];
 
     seccomp = true;
+
+    try_socket = "/tmp/.X11-unix/X0";
+    socket_abstract = true;
+    socket_pathname = false;
   };
 }
diff --git a/test/sandbox/case/pdlike.nix b/test/sandbox/case/pdlike.nix
index da6753f8..ea341478 100644
--- a/test/sandbox/case/pdlike.nix
+++ b/test/sandbox/case/pdlike.nix
@@ -264,7 +264,7 @@ in
     seccomp = true;
 
     try_socket = "/tmp/.X11-unix/X0";
-    socket_abstract = true;
+    socket_abstract = false;
     socket_pathname = false;
   };
 }
diff --git a/test/sandbox/case/preset.nix b/test/sandbox/case/preset.nix
index be376a17..41f612f8 100644
--- a/test/sandbox/case/preset.nix
+++ b/test/sandbox/case/preset.nix
@@ -262,7 +262,7 @@ in
     seccomp = true;
 
     try_socket = "/tmp/.X11-unix/X0";
-    socket_abstract = true;
+    socket_abstract = false;
     socket_pathname = false;
   };
 }
diff --git a/test/sandbox/case/tty.nix b/test/sandbox/case/tty.nix
index ff7bf035..d95cbb83 100644
--- a/test/sandbox/case/tty.nix
+++ b/test/sandbox/case/tty.nix
@@ -275,7 +275,7 @@ in
     seccomp = true;
 
     try_socket = "/tmp/.X11-unix/X0";
-    socket_abstract = true;
+    socket_abstract = false;
     socket_pathname = true;
   };
 }