container: binfmt registration

This arranges for binfmt entries to be registered for the container.

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

container/binfmt.go

@@ -0,0 +1,46 @@
+package container
+
+import (
+	"strings"
+	"unsafe"
+
+	"hakurei.app/check"
+)
+
+// escapeBinfmt escapes magic/mask sequences in a [BinfmtEntry].
+func escapeBinfmt(buf *strings.Builder, s string) string {
+	const lowerhex = "0123456789abcdef"
+
+	buf.Reset()
+	for _, c := range unsafe.Slice(unsafe.StringData(s), len(s)) {
+		switch c {
+		case 0, '\\', ':':
+			buf.WriteString(`\x`)
+			buf.WriteByte(lowerhex[c>>4])
+			buf.WriteByte(lowerhex[c&0xf])
+
+		default:
+			buf.WriteByte(c)
+		}
+	}
+	return buf.String()
+}
+
+// BinfmtEntry is an entry to be registered by the init process.
+type BinfmtEntry struct {
+	// The offset of the magic/mask in the file, counted in bytes.
+	Offset byte
+	// The byte sequence binfmt_misc is matching for.
+	Magic string
+	// An (optional, defaults to all 0xff) mask.
+	Mask string
+	// The program that should be invoked with the binary as first argument.
+	Interpreter *check.Absolute
+}
+
+// Valid returns whether e can be registered into the kernel.
+func (e *BinfmtEntry) Valid() bool {
+	return e != nil &&
+		int(e.Offset)+max(len(e.Magic), len(e.Mask)) < 128 &&
+		e.Interpreter != nil && len(e.Interpreter.String()) < 128
+}

container/binfmt_test.go

@@ -0,0 +1,62 @@
+package container
+
+import (
+	"strings"
+	"testing"
+
+	"hakurei.app/fhs"
+)
+
+func TestEscapeBinfmt(t *testing.T) {
+	t.Parallel()
+
+	testCases := []struct {
+		name  string
+		magic string
+		want  string
+	}{
+		{"packed DOS applications", "\x0eDEX", "\x0eDEX"},
+
+		{"riscv64 magic",
+			"\x7fELF\x02\x01\x01\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02\x00\xf3\x00",
+			"\x7fELF\x02\x01\x01\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\x02\\x00\xf3\\x00"},
+		{"riscv64 mask",
+			"\xff\xff\xff\xff\xff\xff\xff\x00\xff\xff\xff\xff\xff\xff\xff\xff\xfe\xff\xff\xff",
+			"\xff\xff\xff\xff\xff\xff\xff\\x00\xff\xff\xff\xff\xff\xff\xff\xff\xfe\xff\xff\xff"},
+	}
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Parallel()
+
+			got := escapeBinfmt(new(strings.Builder), tc.magic)
+			if got != tc.want {
+				t.Errorf("escapeBinfmt: %q, want %q", got, tc.want)
+			}
+		})
+	}
+}
+
+func TestBinfmtEntry(t *testing.T) {
+	t.Parallel()
+
+	testCases := []struct {
+		name  string
+		e     BinfmtEntry
+		valid bool
+	}{
+		{"zero", BinfmtEntry{}, false},
+		{"large offset", BinfmtEntry{Offset: 128}, false},
+		{"long magic", BinfmtEntry{Magic: strings.Repeat("\x00", 128)}, false},
+		{"long mask", BinfmtEntry{Mask: strings.Repeat("\x00", 128)}, false},
+		{"valid", BinfmtEntry{Interpreter: fhs.AbsRoot}, true},
+	}
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Parallel()
+
+			if tc.e.Valid() != tc.valid {
+				t.Errorf("Valid: %v", !tc.valid)
+			}
+		})
+	}
+}

container/container.go

@@ -100,6 +100,11 @@ type (
 		Gid int
 		// Hostname value in UTS namespace.
 		Hostname string
+		// Register binfmt_misc entries.
+		Binfmt []BinfmtEntry
+		// Alternative pathname to attach binfmt_misc filesystem. The zero value
+		// requires [FstypeProc] to be made available at [fhs.Proc].
+		BinfmtPath *check.Absolute
 		// Sequential container setup ops.
 		*Ops
 
@@ -219,6 +224,9 @@ func (p *Container) Start() error {
 	if p.cmd.Process != nil {
 		return errors.New("container: already started")
 	}
+	if !p.InitAsRoot && len(p.Binfmt) > 0 {
+		return errors.New("container: init as root required, but not enabled")
+	}
 
 	if err := ensureCloseOnExec(); err != nil {
 		return err

container/init.go

@@ -11,11 +11,13 @@ import (
 	"path/filepath"
 	"slices"
 	"strconv"
+	"strings"
 	"sync"
 	"sync/atomic"
 	. "syscall"
 	"time"
 
+	"hakurei.app/check"
 	"hakurei.app/container/seccomp"
 	"hakurei.app/ext"
 	"hakurei.app/fhs"
@@ -240,6 +242,16 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 		k.fatalf(msg, "cannot enter intermediate host path: %v", err)
 	}
 
+	if len(param.Binfmt) > 0 {
+		for i, e := range param.Binfmt {
+			if pathname, err := k.evalSymlinks(e.Interpreter.String()); err != nil {
+				k.fatal(msg, err)
+			} else if param.Binfmt[i].Interpreter, err = check.NewAbs(pathname); err != nil {
+				k.fatal(msg, err)
+			}
+		}
+	}
+
 	/* early is called right before pivot_root into intermediate root;
 	this step is mostly for gathering information that would otherwise be
 	difficult to obtain via library functions after pivot_root, and
@@ -295,6 +307,48 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 		}
 	}
 
+	if len(param.Binfmt) > 0 {
+		const interpreter = "/interpreter"
+
+		if param.BinfmtPath == nil {
+			param.BinfmtPath = fhs.AbsProcSys.Append("fs/binfmt_misc")
+		}
+		binfmt := sysrootPath + param.BinfmtPath.String()
+		if err := k.mkdirAll(binfmt, 0); err != nil {
+			k.fatal(msg, err)
+		}
+		if err := k.mount(
+			SourceBinfmtMisc,
+			binfmt,
+			FstypeBinfmtMisc,
+			MS_NOSUID|MS_NOEXEC|MS_NODEV,
+			zeroString,
+		); err != nil {
+			k.fatal(msg, err)
+		}
+
+		var buf strings.Builder
+		buf.Grow(1920)
+
+		register := binfmt + "/register"
+		for i, e := range param.Binfmt {
+			if err := k.symlink(hostPath+e.Interpreter.String(), interpreter); err != nil {
+				k.fatal(msg, err)
+			} else if err = k.writeFile(register, []byte(":"+
+				strconv.Itoa(i)+":"+
+				"M:"+
+				strconv.Itoa(int(e.Offset))+":"+
+				escapeBinfmt(&buf, e.Magic)+":"+
+				escapeBinfmt(&buf, e.Mask)+":"+
+				interpreter+":"+
+				"F"), 0); err != nil {
+				k.fatal(msg, err)
+			} else if err = k.remove(interpreter); err != nil {
+				k.fatal(msg, err)
+			}
+		}
+	}
+
 	// setup requiring host root complete at this point
 	if err := k.mount(hostDir, hostDir, zeroString, MS_SILENT|MS_REC|MS_PRIVATE, zeroString); err != nil {
 		k.fatalf(msg, "cannot make host root rprivate: %v", optionalErrorUnwrap(err))

container/mount.go

@@ -40,6 +40,9 @@ const (
 	// SourceMqueue is used when mounting mqueue.
 	// Note that any source value is allowed when fstype is [FstypeMqueue].
 	SourceMqueue = "mqueue"
+	// SourceBinfmtMisc is used when mounting binfmt_misc.
+	// Note that any source value is allowed when fstype is [SourceBinfmtMisc].
+	SourceBinfmtMisc = "binfmt_misc"
 	// SourceOverlay is used when mounting overlay.
 	// Note that any source value is allowed when fstype is [FstypeOverlay].
 	SourceOverlay = "overlay"
@@ -70,6 +73,9 @@ const (
 	// FstypeMqueue represents the mqueue pseudo-filesystem.
 	// This filesystem type is usually mounted on /dev/mqueue.
 	FstypeMqueue = "mqueue"
+	// FstypeBinfmtMisc represents the binfmt_misc pseudo-filesystem.
+	// This filesystem type is usually mounted on /proc/sys/fs/binfmt_misc.
+	FstypeBinfmtMisc = "binfmt_misc"
 	// FstypeOverlay represents the overlay pseudo-filesystem.
 	// This filesystem type can be mounted anywhere in the container filesystem.
 	FstypeOverlay = "overlay"

fhs/abs.go

@@ -42,6 +42,8 @@ var (
 	AbsDevShm = unsafeAbs(DevShm)
 	// AbsProc is [Proc] as [check.Absolute].
 	AbsProc = unsafeAbs(Proc)
+	// AbsProcSys is [ProcSys] as [check.Absolute].
+	AbsProcSys = unsafeAbs(ProcSys)
 	// AbsProcSelfExe is [ProcSelfExe] as [check.Absolute].
 	AbsProcSelfExe = unsafeAbs(ProcSelfExe)
 	// AbsSys is [Sys] as [check.Absolute].