sandbox: move out of internal

Signed-off-by: Ophestra <cat@gensokyo.uk>
2025-03-17 02:55:36 +09:00
parent 9ce4706a07
commit 24618ab9a1
37 changed files with 24 additions and 24 deletions
--- a/sandbox/const.go
+++ b/sandbox/const.go
@@ -0,0 +1,6 @@
+package sandbox
+
+const (
+	PR_SET_NO_NEW_PRIVS = 0x26
+	CAP_SYS_ADMIN       = 0x15
+)
--- a/sandbox/container.go
+++ b/sandbox/container.go
@@ -0,0 +1,232 @@
+package sandbox
+
+import (
+	"context"
+	"encoding/gob"
+	"errors"
+	"fmt"
+	"io"
+	"os"
+	"os/exec"
+	"path"
+	"strconv"
+	"syscall"
+	"time"
+
+	"git.gensokyo.uk/security/fortify/sandbox/seccomp"
+)
+
+type HardeningFlags uintptr
+
+const (
+	FSyscallCompat HardeningFlags = 1 << iota
+	FAllowDevel
+	FAllowUserns
+	FAllowTTY
+	FAllowNet
+)
+
+func (flags HardeningFlags) seccomp(opts seccomp.SyscallOpts) seccomp.SyscallOpts {
+	if flags&FSyscallCompat == 0 {
+		opts |= seccomp.FlagExt
+	}
+	if flags&FAllowDevel == 0 {
+		opts |= seccomp.FlagDenyDevel
+	}
+	if flags&FAllowUserns == 0 {
+		opts |= seccomp.FlagDenyNS
+	}
+	if flags&FAllowTTY == 0 {
+		opts |= seccomp.FlagDenyTTY
+	}
+	return opts
+}
+
+type (
+	// Container represents a container environment being prepared or run.
+	// None of [Container] methods are safe for concurrent use.
+	Container struct {
+		// Name of initial process in the container.
+		name string
+		// Cgroup fd, nil to disable.
+		Cgroup *int
+		// ExtraFiles passed through to initial process in the container,
+		// with behaviour identical to its [exec.Cmd] counterpart.
+		ExtraFiles []*os.File
+
+		InitParams
+		// Custom [exec.Cmd] initialisation function.
+		CommandContext func(ctx context.Context) (cmd *exec.Cmd)
+
+		// param encoder for shim and init
+		setup *gob.Encoder
+		// cancels cmd
+		cancel context.CancelFunc
+
+		Stdin  io.Reader
+		Stdout io.Writer
+		Stderr io.Writer
+
+		Cancel    func() error
+		WaitDelay time.Duration
+
+		cmd *exec.Cmd
+		ctx context.Context
+	}
+
+	InitParams struct {
+		// Working directory in the container.
+		Dir string
+		// Initial process environment.
+		Env []string
+		// Absolute path of initial process in the container. Overrides name.
+		Path string
+		// Initial process argv.
+		Args []string
+
+		// Mapped Uid in user namespace.
+		Uid int
+		// Mapped Gid in user namespace.
+		Gid int
+		// Hostname value in UTS namespace.
+		Hostname string
+		// Sequential container setup ops.
+		*Ops
+		// Extra seccomp options.
+		Seccomp seccomp.SyscallOpts
+
+		Flags HardeningFlags
+	}
+
+	Ops []Op
+	Op  interface {
+		apply(params *InitParams) error
+
+		Is(op Op) bool
+		fmt.Stringer
+	}
+)
+
+func (p *Container) Start() error {
+	if p.cmd != nil {
+		return errors.New("sandbox: already started")
+	}
+	if p.Ops == nil || len(*p.Ops) == 0 {
+		return errors.New("sandbox: starting an empty container")
+	}
+
+	ctx, cancel := context.WithCancel(p.ctx)
+	p.cancel = cancel
+
+	var cloneFlags uintptr = syscall.CLONE_NEWIPC |
+		syscall.CLONE_NEWUTS |
+		syscall.CLONE_NEWCGROUP
+	if p.Flags&FAllowNet == 0 {
+		cloneFlags |= syscall.CLONE_NEWNET
+	}
+
+	// map to overflow id to work around ownership checks
+	if p.Uid < 1 {
+		p.Uid = OverflowUid()
+	}
+	if p.Gid < 1 {
+		p.Gid = OverflowGid()
+	}
+
+	if p.CommandContext != nil {
+		p.cmd = p.CommandContext(ctx)
+	} else {
+		p.cmd = exec.CommandContext(ctx, MustExecutable())
+		p.cmd.Args = []string{"init"}
+	}
+
+	p.cmd.Stdin, p.cmd.Stdout, p.cmd.Stderr = p.Stdin, p.Stdout, p.Stderr
+	p.cmd.Cancel, p.cmd.WaitDelay = p.Cancel, p.WaitDelay
+	p.cmd.Dir = "/"
+	p.cmd.SysProcAttr = &syscall.SysProcAttr{
+		Setsid:    p.Flags&FAllowTTY == 0,
+		Pdeathsig: syscall.SIGKILL,
+
+		Cloneflags: cloneFlags |
+			syscall.CLONE_NEWUSER |
+			syscall.CLONE_NEWPID |
+			syscall.CLONE_NEWNS,
+
+		// remain privileged for setup
+		AmbientCaps: []uintptr{CAP_SYS_ADMIN},
+
+		UseCgroupFD: p.Cgroup != nil,
+	}
+	if p.cmd.SysProcAttr.UseCgroupFD {
+		p.cmd.SysProcAttr.CgroupFD = *p.Cgroup
+	}
+
+	// place setup pipe before user supplied extra files, this is later restored by init
+	if fd, e, err := Setup(&p.cmd.ExtraFiles); err != nil {
+		return wrapErrSuffix(err,
+			"cannot create shim setup pipe:")
+	} else {
+		p.setup = e
+		p.cmd.Env = []string{setupEnv + "=" + strconv.Itoa(fd)}
+	}
+	p.cmd.ExtraFiles = append(p.cmd.ExtraFiles, p.ExtraFiles...)
+
+	msg.Verbose("starting container init")
+	if err := p.cmd.Start(); err != nil {
+		return msg.WrapErr(err, err.Error())
+	}
+	return nil
+}
+
+func (p *Container) Serve() error {
+	if p.setup == nil {
+		panic("invalid serve")
+	}
+
+	if p.Path != "" && !path.IsAbs(p.Path) {
+		return msg.WrapErr(syscall.EINVAL,
+			fmt.Sprintf("invalid executable path %q", p.Path))
+	}
+
+	if p.Path == "" {
+		if p.name == "" {
+			p.Path = os.Getenv("SHELL")
+			if !path.IsAbs(p.Path) {
+				return msg.WrapErr(syscall.EBADE,
+					"no command specified and $SHELL is invalid")
+			}
+			p.name = path.Base(p.Path)
+		} else if path.IsAbs(p.name) {
+			p.Path = p.name
+		} else if v, err := exec.LookPath(p.name); err != nil {
+			return msg.WrapErr(err, err.Error())
+		} else {
+			p.Path = v
+		}
+	}
+
+	setup := p.setup
+	p.setup = nil
+	return setup.Encode(
+		&initParams{
+			p.InitParams,
+			syscall.Getuid(),
+			syscall.Getgid(),
+			len(p.ExtraFiles),
+			msg.IsVerbose(),
+		},
+	)
+}
+
+func (p *Container) Wait() error { defer p.cancel(); return p.cmd.Wait() }
+
+func (p *Container) String() string {
+	return fmt.Sprintf("argv: %q, flags: %#x, seccomp: %#x",
+		p.Args, p.Flags, int(p.Flags.seccomp(p.Seccomp)))
+}
+
+func New(ctx context.Context, name string, args ...string) *Container {
+	return &Container{name: name, ctx: ctx,
+		InitParams: InitParams{Args: append([]string{name}, args...), Dir: "/", Ops: new(Ops)},
+	}
+}
--- a/sandbox/container_test.go
+++ b/sandbox/container_test.go
@@ -0,0 +1,182 @@
+package sandbox_test
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"log"
+	"os"
+	"os/exec"
+	"syscall"
+	"testing"
+	"time"
+
+	"git.gensokyo.uk/security/fortify/fst"
+	"git.gensokyo.uk/security/fortify/internal"
+	"git.gensokyo.uk/security/fortify/internal/fmsg"
+	"git.gensokyo.uk/security/fortify/ldd"
+	"git.gensokyo.uk/security/fortify/sandbox"
+	"git.gensokyo.uk/security/fortify/sandbox/seccomp"
+	check "git.gensokyo.uk/security/fortify/test/sandbox"
+)
+
+func TestContainer(t *testing.T) {
+	{
+		oldVerbose := fmsg.Load()
+		oldOutput := sandbox.GetOutput()
+		internal.InstallFmsg(true)
+		t.Cleanup(func() { fmsg.Store(oldVerbose) })
+		t.Cleanup(func() { sandbox.SetOutput(oldOutput) })
+	}
+
+	testCases := []struct {
+		name  string
+		flags sandbox.HardeningFlags
+		ops   *sandbox.Ops
+		mnt   []*check.Mntent
+		host  string
+	}{
+		{"minimal", 0, new(sandbox.Ops), nil, "test-minimal"},
+		{"allow", sandbox.FAllowUserns | sandbox.FAllowNet | sandbox.FAllowTTY,
+			new(sandbox.Ops), nil, "test-minimal"},
+		{"tmpfs", 0,
+			new(sandbox.Ops).
+				Tmpfs(fst.Tmp, 0, 0755),
+			[]*check.Mntent{
+				{FSName: "tmpfs", Dir: fst.Tmp, Type: "tmpfs", Opts: "\x00"},
+			}, "test-tmpfs"},
+		{"dev", sandbox.FAllowTTY, // go test output is not a tty
+			new(sandbox.Ops).
+				Dev("/dev"),
+			[]*check.Mntent{
+				{FSName: "devtmpfs", Dir: "/dev", Type: "tmpfs", Opts: "\x00"},
+				{FSName: "devtmpfs", Dir: "/dev/null", Type: "devtmpfs", Opts: "\x00", Freq: -1, Passno: -1},
+				{FSName: "devtmpfs", Dir: "/dev/zero", Type: "devtmpfs", Opts: "\x00", Freq: -1, Passno: -1},
+				{FSName: "devtmpfs", Dir: "/dev/full", Type: "devtmpfs", Opts: "\x00", Freq: -1, Passno: -1},
+				{FSName: "devtmpfs", Dir: "/dev/random", Type: "devtmpfs", Opts: "\x00", Freq: -1, Passno: -1},
+				{FSName: "devtmpfs", Dir: "/dev/urandom", Type: "devtmpfs", Opts: "\x00", Freq: -1, Passno: -1},
+				{FSName: "devtmpfs", Dir: "/dev/tty", Type: "devtmpfs", Opts: "\x00", Freq: -1, Passno: -1},
+				{FSName: "devpts", Dir: "/dev/pts", Type: "devpts", Opts: "rw,nosuid,noexec,relatime,mode=620,ptmxmode=666", Freq: 0, Passno: 0},
+			}, ""},
+	}
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+			defer cancel()
+
+			container := sandbox.New(ctx, os.Args[0], "-test.v",
+				"-test.run=TestHelperCheckContainer", "--", "check", tc.host)
+			container.Uid = 1000
+			container.Gid = 100
+			container.Hostname = tc.host
+			container.CommandContext = commandContext
+			container.Flags |= tc.flags
+			container.Stdout, container.Stderr = os.Stdout, os.Stderr
+			container.Ops = tc.ops
+			if container.Args[5] == "" {
+				if name, err := os.Hostname(); err != nil {
+					t.Fatalf("cannot get hostname: %v", err)
+				} else {
+					container.Args[5] = name
+				}
+			}
+
+			container.
+				Tmpfs("/tmp", 0, 0755).
+				Bind(os.Args[0], os.Args[0], 0)
+			// in case test has cgo enabled
+			var libPaths []string
+			if entries, err := ldd.ExecFilter(ctx,
+				commandContext,
+				func(v []byte) []byte {
+					return bytes.SplitN(v, []byte("TestHelperInit\n"), 2)[1]
+				}, os.Args[0]); err != nil {
+				log.Fatalf("ldd: %v", err)
+			} else {
+				libPaths = ldd.Path(entries)
+			}
+			for _, name := range libPaths {
+				container.Bind(name, name, 0)
+			}
+
+			mnt := make([]*check.Mntent, 0, 3+len(libPaths))
+			mnt = append(mnt, &check.Mntent{FSName: "rootfs", Dir: "/", Type: "tmpfs", Opts: "host_passthrough"})
+			mnt = append(mnt, tc.mnt...)
+			mnt = append(mnt,
+				&check.Mntent{FSName: "tmpfs", Dir: "/tmp", Type: "tmpfs", Opts: "host_passthrough"},
+				&check.Mntent{FSName: "\x00", Dir: os.Args[0], Type: "\x00", Opts: "\x00"})
+			for _, name := range libPaths {
+				mnt = append(mnt, &check.Mntent{FSName: "\x00", Dir: name, Type: "\x00", Opts: "\x00", Freq: -1, Passno: -1})
+			}
+			mnt = append(mnt, &check.Mntent{FSName: "proc", Dir: "/proc", Type: "proc", Opts: "rw,nosuid,nodev,noexec,relatime"})
+			mntentWant := new(bytes.Buffer)
+			if err := json.NewEncoder(mntentWant).Encode(mnt); err != nil {
+				t.Fatalf("cannot serialise mntent: %v", err)
+			}
+			container.Stdin = mntentWant
+
+			// needs /proc to check mntent
+			container.Proc("/proc")
+
+			if err := container.Start(); err != nil {
+				fmsg.PrintBaseError(err, "start:")
+				t.Fatalf("cannot start container: %v", err)
+			} else if err = container.Serve(); err != nil {
+				fmsg.PrintBaseError(err, "serve:")
+				t.Errorf("cannot serve setup params: %v", err)
+			}
+			if err := container.Wait(); err != nil {
+				fmsg.PrintBaseError(err, "wait:")
+				t.Fatalf("wait: %v", err)
+			}
+		})
+	}
+}
+
+func TestContainerString(t *testing.T) {
+	container := sandbox.New(context.TODO(), "ldd", "/usr/bin/env")
+	container.Flags |= sandbox.FAllowDevel
+	container.Seccomp |= seccomp.FlagMultiarch
+	want := `argv: ["ldd" "/usr/bin/env"], flags: 0x2, seccomp: 0x2e`
+	if got := container.String(); got != want {
+		t.Errorf("String: %s, want %s", got, want)
+	}
+}
+
+func TestHelperInit(t *testing.T) {
+	if len(os.Args) != 5 || os.Args[4] != "init" {
+		return
+	}
+	sandbox.SetOutput(fmsg.Output{})
+	sandbox.Init(fmsg.Prepare, internal.InstallFmsg)
+}
+
+func TestHelperCheckContainer(t *testing.T) {
+	if len(os.Args) != 6 || os.Args[4] != "check" {
+		return
+	}
+
+	t.Run("user", func(t *testing.T) {
+		if uid := syscall.Getuid(); uid != 1000 {
+			t.Errorf("Getuid: %d, want 1000", uid)
+		}
+		if gid := syscall.Getgid(); gid != 100 {
+			t.Errorf("Getgid: %d, want 100", gid)
+		}
+	})
+	t.Run("hostname", func(t *testing.T) {
+		if name, err := os.Hostname(); err != nil {
+			t.Fatalf("cannot get hostname: %v", err)
+		} else if name != os.Args[5] {
+			t.Errorf("Hostname: %q, want %q", name, os.Args[5])
+		}
+	})
+	t.Run("seccomp", func(t *testing.T) { check.MustAssertSeccomp() })
+	t.Run("mntent", func(t *testing.T) { check.MustAssertMounts("", "/proc/mounts", "/proc/self/fd/0") })
+}
+
+func commandContext(ctx context.Context) *exec.Cmd {
+	return exec.CommandContext(ctx, os.Args[0], "-test.v",
+		"-test.run=TestHelperInit", "--", "init")
+}
--- a/sandbox/executable.go
+++ b/sandbox/executable.go
@@ -0,0 +1,26 @@
+package sandbox
+
+import (
+	"log"
+	"os"
+	"sync"
+)
+
+var (
+	executable     string
+	executableOnce sync.Once
+)
+
+func copyExecutable() {
+	if name, err := os.Executable(); err != nil {
+		msg.BeforeExit()
+		log.Fatalf("cannot read executable path: %v", err)
+	} else {
+		executable = name
+	}
+}
+
+func MustExecutable() string {
+	executableOnce.Do(copyExecutable)
+	return executable
+}
--- a/sandbox/executable_test.go
+++ b/sandbox/executable_test.go
@@ -0,0 +1,17 @@
+package sandbox_test
+
+import (
+	"os"
+	"testing"
+
+	"git.gensokyo.uk/security/fortify/sandbox"
+)
+
+func TestExecutable(t *testing.T) {
+	for i := 0; i < 16; i++ {
+		if got := sandbox.MustExecutable(); got != os.Args[0] {
+			t.Errorf("MustExecutable: %q, want %q",
+				got, os.Args[0])
+		}
+	}
+}
--- a/sandbox/init.go
+++ b/sandbox/init.go
@@ -0,0 +1,341 @@
+package sandbox
+
+import (
+	"errors"
+	"fmt"
+	"log"
+	"os"
+	"os/exec"
+	"os/signal"
+	"path"
+	"runtime"
+	"strconv"
+	"syscall"
+	"time"
+
+	"git.gensokyo.uk/security/fortify/sandbox/seccomp"
+)
+
+const (
+	// time to wait for linger processes after death of initial process
+	residualProcessTimeout = 5 * time.Second
+
+	// intermediate tmpfs mount point
+	basePath = "/tmp"
+
+	// setup params file descriptor
+	setupEnv = "FORTIFY_SETUP"
+)
+
+type initParams struct {
+	InitParams
+
+	HostUid, HostGid int
+	// extra files count
+	Count int
+	// verbosity pass through
+	Verbose bool
+}
+
+func Init(prepare func(prefix string), setVerbose func(verbose bool)) {
+	runtime.LockOSThread()
+	prepare("init")
+
+	if os.Getpid() != 1 {
+		log.Fatal("this process must run as pid 1")
+	}
+
+	/*
+		receive setup payload
+	*/
+
+	var (
+		params      initParams
+		closeSetup  func() error
+		setupFile   *os.File
+		offsetSetup int
+	)
+	if f, err := Receive(setupEnv, &params, &setupFile); err != nil {
+		if errors.Is(err, ErrInvalid) {
+			log.Fatal("invalid setup descriptor")
+		}
+		if errors.Is(err, ErrNotSet) {
+			log.Fatal("FORTIFY_SETUP not set")
+		}
+
+		log.Fatalf("cannot decode init setup payload: %v", err)
+	} else {
+		if params.Ops == nil {
+			log.Fatal("invalid setup parameters")
+		}
+
+		setVerbose(params.Verbose)
+		msg.Verbose("received setup parameters")
+		closeSetup = f
+		offsetSetup = int(setupFile.Fd() + 1)
+	}
+
+	// write uid/gid map here so parent does not need to set dumpable
+	if err := SetDumpable(SUID_DUMP_USER); err != nil {
+		log.Fatalf("cannot set SUID_DUMP_USER: %s", err)
+	}
+	if err := os.WriteFile("/proc/self/uid_map",
+		append([]byte{}, strconv.Itoa(params.Uid)+" "+strconv.Itoa(params.HostUid)+" 1\n"...),
+		0); err != nil {
+		log.Fatalf("%v", err)
+	}
+	if err := os.WriteFile("/proc/self/setgroups",
+		[]byte("deny\n"),
+		0); err != nil && !os.IsNotExist(err) {
+		log.Fatalf("%v", err)
+	}
+	if err := os.WriteFile("/proc/self/gid_map",
+		append([]byte{}, strconv.Itoa(params.Gid)+" "+strconv.Itoa(params.HostGid)+" 1\n"...),
+		0); err != nil {
+		log.Fatalf("%v", err)
+	}
+	if err := SetDumpable(SUID_DUMP_DISABLE); err != nil {
+		log.Fatalf("cannot set SUID_DUMP_DISABLE: %s", err)
+	}
+
+	if params.Hostname != "" {
+		if err := syscall.Sethostname([]byte(params.Hostname)); err != nil {
+			log.Fatalf("cannot set hostname: %v", err)
+		}
+	}
+
+	/*
+		set up mount points from intermediate root
+	*/
+
+	if err := syscall.Mount("", "/", "",
+		syscall.MS_SILENT|syscall.MS_SLAVE|syscall.MS_REC,
+		""); err != nil {
+		log.Fatalf("cannot make / rslave: %v", err)
+	}
+
+	if err := syscall.Mount("rootfs", basePath, "tmpfs",
+		syscall.MS_NODEV|syscall.MS_NOSUID,
+		""); err != nil {
+		log.Fatalf("cannot mount intermediate root: %v", err)
+	}
+	if err := os.Chdir(basePath); err != nil {
+		log.Fatalf("cannot enter base path: %v", err)
+	}
+
+	if err := os.Mkdir(sysrootDir, 0755); err != nil {
+		log.Fatalf("%v", err)
+	}
+	if err := syscall.Mount(sysrootDir, sysrootDir, "",
+		syscall.MS_SILENT|syscall.MS_MGC_VAL|syscall.MS_BIND|syscall.MS_REC,
+		""); err != nil {
+		log.Fatalf("cannot bind sysroot: %v", err)
+	}
+
+	if err := os.Mkdir(hostDir, 0755); err != nil {
+		log.Fatalf("%v", err)
+	}
+	if err := syscall.PivotRoot(basePath, hostDir); err != nil {
+		log.Fatalf("cannot pivot into intermediate root: %v", err)
+	}
+	if err := os.Chdir("/"); err != nil {
+		log.Fatalf("%v", err)
+	}
+
+	for i, op := range *params.Ops {
+		msg.Verbosef("mounting %s", op)
+		if err := op.apply(&params.InitParams); err != nil {
+			msg.PrintBaseErr(err,
+				fmt.Sprintf("cannot apply op %d:", i))
+			msg.BeforeExit()
+			os.Exit(1)
+		}
+	}
+
+	/*
+		pivot to sysroot
+	*/
+
+	if err := syscall.Mount(hostDir, hostDir, "",
+		syscall.MS_SILENT|syscall.MS_REC|syscall.MS_PRIVATE,
+		""); err != nil {
+		log.Fatalf("cannot make host root rprivate: %v", err)
+	}
+	if err := syscall.Unmount(hostDir, syscall.MNT_DETACH); err != nil {
+		log.Fatalf("cannot unmount host root: %v", err)
+	}
+
+	{
+		var fd int
+		if err := IgnoringEINTR(func() (err error) {
+			fd, err = syscall.Open("/", syscall.O_DIRECTORY|syscall.O_RDONLY, 0)
+			return
+		}); err != nil {
+			log.Fatalf("cannot open intermediate root: %v", err)
+		}
+		if err := os.Chdir(sysrootPath); err != nil {
+			log.Fatalf("%v", err)
+		}
+
+		if err := syscall.PivotRoot(".", "."); err != nil {
+			log.Fatalf("cannot pivot into sysroot: %v", err)
+		}
+		if err := syscall.Fchdir(fd); err != nil {
+			log.Fatalf("cannot re-enter intermediate root: %v", err)
+		}
+		if err := syscall.Unmount(".", syscall.MNT_DETACH); err != nil {
+			log.Fatalf("cannot unmount intemediate root: %v", err)
+		}
+		if err := os.Chdir("/"); err != nil {
+			log.Fatalf("%v", err)
+		}
+
+		if err := syscall.Close(fd); err != nil {
+			log.Fatalf("cannot close intermediate root: %v", err)
+		}
+	}
+
+	/*
+		load seccomp filter
+	*/
+
+	if _, _, err := syscall.Syscall(PR_SET_NO_NEW_PRIVS, 1, 0, 0); err != 0 {
+		log.Fatalf("prctl(PR_SET_NO_NEW_PRIVS): %v", err)
+	}
+	if err := seccomp.Load(params.Flags.seccomp(params.Seccomp)); err != nil {
+		log.Fatalf("cannot load syscall filter: %v", err)
+	}
+
+	/* at this point CAP_SYS_ADMIN can be dropped, however it is kept for now as it does not increase attack surface */
+
+	/*
+		pass through extra files
+	*/
+
+	extraFiles := make([]*os.File, params.Count)
+	for i := range extraFiles {
+		extraFiles[i] = os.NewFile(uintptr(offsetSetup+i), "extra file "+strconv.Itoa(i))
+	}
+
+	/*
+		prepare initial process
+	*/
+
+	cmd := exec.Command(params.Path)
+	cmd.Stdin, cmd.Stdout, cmd.Stderr = os.Stdin, os.Stdout, os.Stderr
+	cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
+	cmd.Args = params.Args
+	cmd.Env = params.Env
+	cmd.ExtraFiles = extraFiles
+	cmd.Dir = params.Dir
+
+	if err := cmd.Start(); err != nil {
+		log.Fatalf("%v", err)
+	}
+	msg.Suspend()
+
+	/*
+		close setup pipe
+	*/
+
+	if err := closeSetup(); err != nil {
+		log.Println("cannot close setup pipe:", err)
+		// not fatal
+	}
+
+	/*
+		perform init duties
+	*/
+
+	sig := make(chan os.Signal, 2)
+	signal.Notify(sig, syscall.SIGINT, syscall.SIGTERM)
+
+	type winfo struct {
+		wpid    int
+		wstatus syscall.WaitStatus
+	}
+	info := make(chan winfo, 1)
+	done := make(chan struct{})
+
+	go func() {
+		var (
+			err     error
+			wpid    = -2
+			wstatus syscall.WaitStatus
+		)
+
+		// keep going until no child process is left
+		for wpid != -1 {
+			if err != nil {
+				break
+			}
+
+			if wpid != -2 {
+				info <- winfo{wpid, wstatus}
+			}
+
+			err = syscall.EINTR
+			for errors.Is(err, syscall.EINTR) {
+				wpid, err = syscall.Wait4(-1, &wstatus, 0, nil)
+			}
+		}
+		if !errors.Is(err, syscall.ECHILD) {
+			log.Println("unexpected wait4 response:", err)
+		}
+
+		close(done)
+	}()
+
+	// closed after residualProcessTimeout has elapsed after initial process death
+	timeout := make(chan struct{})
+
+	r := 2
+	for {
+		select {
+		case s := <-sig:
+			if msg.Resume() {
+				msg.Verbosef("terminating on %s after process start", s.String())
+			} else {
+				msg.Verbosef("terminating on %s", s.String())
+			}
+			msg.BeforeExit()
+			os.Exit(0)
+		case w := <-info:
+			if w.wpid == cmd.Process.Pid {
+				// initial process exited, output is most likely available again
+				msg.Resume()
+
+				switch {
+				case w.wstatus.Exited():
+					r = w.wstatus.ExitStatus()
+				case w.wstatus.Signaled():
+					r = 128 + int(w.wstatus.Signal())
+				default:
+					r = 255
+				}
+
+				go func() {
+					time.Sleep(residualProcessTimeout)
+					close(timeout)
+				}()
+			}
+		case <-done:
+			msg.BeforeExit()
+			os.Exit(r)
+		case <-timeout:
+			log.Println("timeout exceeded waiting for lingering processes")
+			msg.BeforeExit()
+			os.Exit(r)
+		}
+	}
+}
+
+// TryArgv0 calls [Init] if the last element of argv0 is "init".
+func TryArgv0(v Msg, prepare func(prefix string), setVerbose func(verbose bool)) {
+	if len(os.Args) > 0 && path.Base(os.Args[0]) == "init" {
+		msg = v
+		Init(prepare, setVerbose)
+		msg.BeforeExit()
+		os.Exit(0)
+	}
+}
--- a/sandbox/mount.go
+++ b/sandbox/mount.go
@@ -0,0 +1,95 @@
+package sandbox
+
+import (
+	"errors"
+	"fmt"
+	"os"
+	"strings"
+	"syscall"
+)
+
+const (
+	BindOptional = 1 << iota
+	BindSource
+	BindRecursive
+	BindWritable
+	BindDevices
+)
+
+func bindMount(src, dest string, flags int) error {
+	target := toSysroot(dest)
+	var source string
+
+	if flags&BindSource == 0 {
+		// this is what bwrap does, so the behaviour is kept for now,
+		// however recursively resolving links might improve user experience
+		if rp, err := realpathHost(src); err != nil {
+			if os.IsNotExist(err) {
+				if flags&BindOptional != 0 {
+					return nil
+				} else {
+					return msg.WrapErr(err,
+						fmt.Sprintf("path %q does not exist", src))
+				}
+			}
+			return msg.WrapErr(err, err.Error())
+		} else {
+			source = toHost(rp)
+		}
+	} else if flags&BindOptional != 0 {
+		return msg.WrapErr(syscall.EINVAL,
+			"flag source excludes optional")
+	} else {
+		source = toHost(src)
+	}
+
+	if fi, err := os.Stat(source); err != nil {
+		return msg.WrapErr(err, err.Error())
+	} else if fi.IsDir() {
+		if err = os.MkdirAll(target, 0755); err != nil {
+			return wrapErrSuffix(err,
+				fmt.Sprintf("cannot create directory %q:", dest))
+		}
+	} else if err = ensureFile(target, 0444); err != nil {
+		if errors.Is(err, syscall.EISDIR) {
+			return msg.WrapErr(err,
+				fmt.Sprintf("path %q is a directory", dest))
+		}
+		return wrapErrSuffix(err,
+			fmt.Sprintf("cannot create %q:", dest))
+	}
+
+	var mf uintptr = syscall.MS_SILENT | syscall.MS_BIND
+	if flags&BindRecursive != 0 {
+		mf |= syscall.MS_REC
+	}
+	if flags&BindWritable == 0 {
+		mf |= syscall.MS_RDONLY
+	}
+	if flags&BindDevices == 0 {
+		mf |= syscall.MS_NODEV
+	}
+	if msg.IsVerbose() {
+		if strings.TrimPrefix(source, hostPath) == strings.TrimPrefix(target, sysrootPath) {
+			msg.Verbosef("resolved %q flags %#x", target, mf)
+		} else {
+			msg.Verbosef("resolved %q on %q flags %#x", source, target, mf)
+		}
+	}
+	return wrapErrSuffix(syscall.Mount(source, target, "", mf, ""),
+		fmt.Sprintf("cannot bind %q on %q:", src, dest))
+}
+
+func mountTmpfs(fsname, name string, size int, perm os.FileMode) error {
+	target := toSysroot(name)
+	if err := os.MkdirAll(target, perm); err != nil {
+		return err
+	}
+	opt := fmt.Sprintf("mode=%#o", perm)
+	if size > 0 {
+		opt += fmt.Sprintf(",size=%d", size)
+	}
+	return wrapErrSuffix(syscall.Mount(fsname, target, "tmpfs",
+		syscall.MS_NOSUID|syscall.MS_NODEV, opt),
+		fmt.Sprintf("cannot mount tmpfs on %q:", name))
+}
--- a/sandbox/msg.go
+++ b/sandbox/msg.go
@@ -0,0 +1,43 @@
+package sandbox
+
+import (
+	"log"
+	"sync/atomic"
+)
+
+type Msg interface {
+	IsVerbose() bool
+	Verbose(v ...any)
+	Verbosef(format string, v ...any)
+	WrapErr(err error, a ...any) error
+	PrintBaseErr(err error, fallback string)
+
+	Suspend()
+	Resume() bool
+
+	BeforeExit()
+}
+
+type DefaultMsg struct{ inactive atomic.Bool }
+
+func (msg *DefaultMsg) IsVerbose() bool { return true }
+func (msg *DefaultMsg) Verbose(v ...any) {
+	if !msg.inactive.Load() {
+		log.Println(v...)
+	}
+}
+func (msg *DefaultMsg) Verbosef(format string, v ...any) {
+	if !msg.inactive.Load() {
+		log.Printf(format, v...)
+	}
+}
+
+func (msg *DefaultMsg) WrapErr(err error, a ...any) error {
+	log.Println(a...)
+	return err
+}
+func (msg *DefaultMsg) PrintBaseErr(err error, fallback string) { log.Println(fallback, err) }
+
+func (msg *DefaultMsg) Suspend()     { msg.inactive.Store(true) }
+func (msg *DefaultMsg) Resume() bool { return msg.inactive.CompareAndSwap(true, false) }
+func (msg *DefaultMsg) BeforeExit()  {}
--- a/sandbox/output.go
+++ b/sandbox/output.go
@@ -0,0 +1,19 @@
+package sandbox
+
+var msg Msg = new(DefaultMsg)
+
+func GetOutput() Msg { return msg }
+func SetOutput(v Msg) {
+	if v == nil {
+		msg = new(DefaultMsg)
+	} else {
+		msg = v
+	}
+}
+
+func wrapErrSuffix(err error, a ...any) error {
+	if err == nil {
+		return nil
+	}
+	return msg.WrapErr(err, append(a, err)...)
+}
--- a/sandbox/overflow.go
+++ b/sandbox/overflow.go
@@ -0,0 +1,37 @@
+package sandbox
+
+import (
+	"bytes"
+	"log"
+	"os"
+	"strconv"
+	"sync"
+)
+
+var (
+	ofUid  int
+	ofGid  int
+	ofOnce sync.Once
+)
+
+const (
+	ofUidPath = "/proc/sys/kernel/overflowuid"
+	ofGidPath = "/proc/sys/kernel/overflowgid"
+)
+
+func mustReadOverflow() {
+	if v, err := os.ReadFile(ofUidPath); err != nil {
+		log.Fatalf("cannot read %q: %v", ofUidPath, err)
+	} else if ofUid, err = strconv.Atoi(string(bytes.TrimSpace(v))); err != nil {
+		log.Fatalf("cannot interpret %q: %v", ofUidPath, err)
+	}
+
+	if v, err := os.ReadFile(ofGidPath); err != nil {
+		log.Fatalf("cannot read %q: %v", ofGidPath, err)
+	} else if ofGid, err = strconv.Atoi(string(bytes.TrimSpace(v))); err != nil {
+		log.Fatalf("cannot interpret %q: %v", ofGidPath, err)
+	}
+}
+
+func OverflowUid() int { ofOnce.Do(mustReadOverflow); return ofUid }
+func OverflowGid() int { ofOnce.Do(mustReadOverflow); return ofGid }
--- a/sandbox/params.go
+++ b/sandbox/params.go
@@ -0,0 +1,47 @@
+package sandbox
+
+import (
+	"encoding/gob"
+	"errors"
+	"os"
+	"strconv"
+)
+
+var (
+	ErrNotSet  = errors.New("environment variable not set")
+	ErrInvalid = errors.New("bad file descriptor")
+)
+
+// Setup appends the read end of a pipe for setup params transmission and returns its fd.
+func Setup(extraFiles *[]*os.File) (int, *gob.Encoder, error) {
+	if r, w, err := os.Pipe(); err != nil {
+		return -1, nil, err
+	} else {
+		fd := 3 + len(*extraFiles)
+		*extraFiles = append(*extraFiles, r)
+		return fd, gob.NewEncoder(w), nil
+	}
+}
+
+// Receive retrieves setup fd from the environment and receives params.
+func Receive(key string, e any, v **os.File) (func() error, error) {
+	var setup *os.File
+
+	if s, ok := os.LookupEnv(key); !ok {
+		return nil, ErrNotSet
+	} else {
+		if fd, err := strconv.Atoi(s); err != nil {
+			return nil, err
+		} else {
+			setup = os.NewFile(uintptr(fd), "setup")
+			if setup == nil {
+				return nil, ErrInvalid
+			}
+			if v != nil {
+				*v = setup
+			}
+		}
+	}
+
+	return setup.Close, gob.NewDecoder(setup).Decode(e)
+}
--- a/sandbox/path.go
+++ b/sandbox/path.go
@@ -0,0 +1,75 @@
+package sandbox
+
+import (
+	"errors"
+	"io/fs"
+	"os"
+	"path"
+	"strings"
+	"syscall"
+)
+
+const (
+	hostPath    = "/" + hostDir
+	hostDir     = "host"
+	sysrootPath = "/" + sysrootDir
+	sysrootDir  = "sysroot"
+)
+
+func toSysroot(name string) string {
+	name = strings.TrimLeftFunc(name, func(r rune) bool { return r == '/' })
+	return path.Join(sysrootPath, name)
+}
+
+func toHost(name string) string {
+	name = strings.TrimLeftFunc(name, func(r rune) bool { return r == '/' })
+	return path.Join(hostPath, name)
+}
+
+func realpathHost(name string) (string, error) {
+	source := toHost(name)
+	rp, err := os.Readlink(source)
+
+	if err != nil {
+		if errors.Is(err, syscall.EINVAL) {
+			// not a symlink
+			return name, nil
+		}
+		return "", err
+	}
+
+	if !path.IsAbs(rp) {
+		return name, nil
+	}
+	msg.Verbosef("path %q resolves to %q", name, rp)
+	return rp, nil
+}
+
+func createFile(name string, perm os.FileMode, content []byte) error {
+	if err := os.MkdirAll(path.Dir(name), 0755); err != nil {
+		return err
+	}
+	f, err := os.OpenFile(name, syscall.O_CREAT|syscall.O_EXCL|syscall.O_WRONLY, perm)
+	if err != nil {
+		return err
+	}
+	if content != nil {
+		_, err = f.Write(content)
+	}
+	return errors.Join(f.Close(), err)
+}
+
+func ensureFile(name string, perm os.FileMode) error {
+	fi, err := os.Stat(name)
+	if err != nil {
+		if !os.IsNotExist(err) {
+			return err
+		}
+		return createFile(name, perm, nil)
+	}
+
+	if mode := fi.Mode(); mode&fs.ModeDir != 0 || mode&fs.ModeSymlink != 0 {
+		err = syscall.EISDIR
+	}
+	return err
+}
--- a/sandbox/seccomp/api.go
+++ b/sandbox/seccomp/api.go
@@ -0,0 +1,71 @@
+package seccomp
+
+import (
+	"context"
+	"errors"
+	"syscall"
+
+	"git.gensokyo.uk/security/fortify/helper/proc"
+)
+
+// New returns an inactive Encoder instance.
+func New(opts SyscallOpts) *Encoder { return &Encoder{newExporter(opts)} }
+
+// Load loads a filter into the kernel.
+func Load(opts SyscallOpts) error { return buildFilter(-1, opts) }
+
+/*
+An Encoder writes a BPF program to an output stream.
+
+Methods of Encoder are not safe for concurrent use.
+
+An Encoder must not be copied after first use.
+*/
+type Encoder struct {
+	*exporter
+}
+
+func (e *Encoder) Read(p []byte) (n int, err error) {
+	if err = e.prepare(); err != nil {
+		return
+	}
+	return e.r.Read(p)
+}
+
+func (e *Encoder) Close() error {
+	if e.r == nil {
+		return syscall.EINVAL
+	}
+
+	// this hangs if the cgo thread fails to exit
+	return errors.Join(e.closeWrite(), <-e.exportErr)
+}
+
+// NewFile returns an instance of exporter implementing [proc.File].
+func NewFile(opts SyscallOpts) proc.File { return &File{opts: opts} }
+
+// File implements [proc.File] and provides access to the read end of exporter pipe.
+type File struct {
+	opts SyscallOpts
+	proc.BaseFile
+}
+
+func (f *File) ErrCount() int { return 2 }
+func (f *File) Fulfill(ctx context.Context, dispatchErr func(error)) error {
+	e := newExporter(f.opts)
+	if err := e.prepare(); err != nil {
+		return err
+	}
+	f.Set(e.r)
+	go func() {
+		select {
+		case err := <-e.exportErr:
+			dispatchErr(nil)
+			dispatchErr(err)
+		case <-ctx.Done():
+			dispatchErr(e.closeWrite())
+			dispatchErr(<-e.exportErr)
+		}
+	}()
+	return nil
+}
--- a/sandbox/seccomp/export.go
+++ b/sandbox/seccomp/export.go
@@ -0,0 +1,58 @@
+package seccomp
+
+import (
+	"os"
+	"runtime"
+	"sync"
+)
+
+type exporter struct {
+	opts SyscallOpts
+	r, w *os.File
+
+	prepareOnce sync.Once
+	prepareErr  error
+	closeOnce   sync.Once
+	closeErr    error
+	exportErr   <-chan error
+}
+
+func (e *exporter) prepare() error {
+	e.prepareOnce.Do(func() {
+		if r, w, err := os.Pipe(); err != nil {
+			e.prepareErr = err
+			return
+		} else {
+			e.r, e.w = r, w
+		}
+
+		ec := make(chan error, 1)
+		go func(fd uintptr) {
+			ec <- buildFilter(int(fd), e.opts)
+			close(ec)
+			_ = e.closeWrite()
+			runtime.KeepAlive(e.w)
+		}(e.w.Fd())
+		e.exportErr = ec
+		runtime.SetFinalizer(e, (*exporter).closeWrite)
+	})
+	return e.prepareErr
+}
+
+func (e *exporter) closeWrite() error {
+	e.closeOnce.Do(func() {
+		if e.w == nil {
+			panic("closeWrite called on invalid exporter")
+		}
+		e.closeErr = e.w.Close()
+
+		// no need for a finalizer anymore
+		runtime.SetFinalizer(e, nil)
+	})
+
+	return e.closeErr
+}
+
+func newExporter(opts SyscallOpts) *exporter {
+	return &exporter{opts: opts}
+}
--- a/sandbox/seccomp/export_test.go
+++ b/sandbox/seccomp/export_test.go
@@ -0,0 +1,139 @@
+package seccomp_test
+
+import (
+	"crypto/sha512"
+	"errors"
+	"io"
+	"slices"
+	"syscall"
+	"testing"
+
+	"git.gensokyo.uk/security/fortify/sandbox/seccomp"
+)
+
+func TestExport(t *testing.T) {
+	testCases := []struct {
+		name    string
+		opts    seccomp.SyscallOpts
+		want    []byte
+		wantErr bool
+	}{
+		{"compat", 0, []byte{
+			0x95, 0xec, 0x69, 0xd0, 0x17, 0x73, 0x3e, 0x07,
+			0x21, 0x60, 0xe0, 0xda, 0x80, 0xfd, 0xeb, 0xec,
+			0xdf, 0x27, 0xae, 0x81, 0x66, 0xf5, 0xe2, 0xa7,
+			0x31, 0x27, 0x0c, 0x98, 0xea, 0x2d, 0x29, 0x46,
+			0xcb, 0x52, 0x31, 0x02, 0x90, 0x63, 0x66, 0x8a,
+			0xf2, 0x15, 0x87, 0x91, 0x55, 0xda, 0x21, 0xac,
+			0xa7, 0x9b, 0x07, 0x0e, 0x04, 0xc0, 0xee, 0x9a,
+			0xcd, 0xf5, 0x8f, 0x55, 0xcf, 0xa8, 0x15, 0xa5,
+		}, false},
+		{"base", seccomp.FlagExt, []byte{
+			0xdc, 0x7f, 0x2e, 0x1c, 0x5e, 0x82, 0x9b, 0x79,
+			0xeb, 0xb7, 0xef, 0xc7, 0x59, 0x15, 0x0f, 0x54,
+			0xa8, 0x3a, 0x75, 0xc8, 0xdf, 0x6f, 0xee, 0x4d,
+			0xce, 0x5d, 0xad, 0xc4, 0x73, 0x6c, 0x58, 0x5d,
+			0x4d, 0xee, 0xbf, 0xeb, 0x3c, 0x79, 0x69, 0xaf,
+			0x3a, 0x07, 0x7e, 0x90, 0xb7, 0x7b, 0xb4, 0x74,
+			0x1d, 0xb0, 0x5d, 0x90, 0x99, 0x7c, 0x86, 0x59,
+			0xb9, 0x58, 0x91, 0x20, 0x6a, 0xc9, 0x95, 0x2d,
+		}, false},
+		{"everything", seccomp.FlagExt |
+			seccomp.FlagDenyNS | seccomp.FlagDenyTTY | seccomp.FlagDenyDevel |
+			seccomp.FlagMultiarch | seccomp.FlagLinux32 | seccomp.FlagCan |
+			seccomp.FlagBluetooth, []byte{
+			0xe9, 0x9d, 0xd3, 0x45, 0xe1, 0x95, 0x41, 0x34,
+			0x73, 0xd3, 0xcb, 0xee, 0x07, 0xb4, 0xed, 0x57,
+			0xb9, 0x08, 0xbf, 0xa8, 0x9e, 0xa2, 0x07, 0x2f,
+			0xe9, 0x34, 0x82, 0x84, 0x7f, 0x50, 0xb5, 0xb7,
+			0x58, 0xda, 0x17, 0xe7, 0x4c, 0xa2, 0xbb, 0xc0,
+			0x08, 0x13, 0xde, 0x49, 0xa2, 0xb9, 0xbf, 0x83,
+			0x4c, 0x02, 0x4e, 0xd4, 0x88, 0x50, 0xbe, 0x69,
+			0xb6, 0x8a, 0x9a, 0x4c, 0x5f, 0x53, 0xa9, 0xdb,
+		}, false},
+		{"strict", seccomp.FlagExt |
+			seccomp.FlagDenyNS | seccomp.FlagDenyTTY | seccomp.FlagDenyDevel, []byte{
+			0xe8, 0x80, 0x29, 0x8d, 0xf2, 0xbd, 0x67, 0x51,
+			0xd0, 0x04, 0x0f, 0xc2, 0x1b, 0xc0, 0xed, 0x4c,
+			0x00, 0xf9, 0x5d, 0xc0, 0xd7, 0xba, 0x50, 0x6c,
+			0x24, 0x4d, 0x8b, 0x8c, 0xf6, 0x86, 0x6d, 0xba,
+			0x8e, 0xf4, 0xa3, 0x32, 0x96, 0xf2, 0x87, 0xb6,
+			0x6c, 0xcc, 0xc1, 0xd7, 0x8e, 0x97, 0x02, 0x65,
+			0x97, 0xf8, 0x4c, 0xc7, 0xde, 0xc1, 0x57, 0x3e,
+			0x14, 0x89, 0x60, 0xfb, 0xd3, 0x5c, 0xd7, 0x35,
+		}, false},
+		{"strict compat", 0 |
+			seccomp.FlagDenyNS | seccomp.FlagDenyTTY | seccomp.FlagDenyDevel, []byte{
+			0x39, 0x87, 0x1b, 0x93, 0xff, 0xaf, 0xc8, 0xb9,
+			0x79, 0xfc, 0xed, 0xc0, 0xb0, 0xc3, 0x7b, 0x9e,
+			0x03, 0x92, 0x2f, 0x5b, 0x02, 0x74, 0x8d, 0xc5,
+			0xc3, 0xc1, 0x7c, 0x92, 0x52, 0x7f, 0x6e, 0x02,
+			0x2e, 0xde, 0x1f, 0x48, 0xbf, 0xf5, 0x92, 0x46,
+			0xea, 0x45, 0x2c, 0x0d, 0x1d, 0xe5, 0x48, 0x27,
+			0x80, 0x8b, 0x1a, 0x6f, 0x84, 0xf3, 0x2b, 0xbd,
+			0xe1, 0xaa, 0x02, 0xae, 0x30, 0xee, 0xdc, 0xfa,
+		}, false},
+	}
+
+	buf := make([]byte, 8)
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			oldF := seccomp.GetOutput()
+			seccomp.SetOutput(t.Log)
+			t.Cleanup(func() { seccomp.SetOutput(oldF) })
+
+			e := seccomp.New(tc.opts)
+			digest := sha512.New()
+
+			if _, err := io.CopyBuffer(digest, e, buf); (err != nil) != tc.wantErr {
+				t.Errorf("Exporter: error = %v, wantErr %v", err, tc.wantErr)
+				return
+			}
+			if err := e.Close(); err != nil {
+				t.Errorf("Close: error = %v", err)
+				return
+			}
+			if got := digest.Sum(nil); slices.Compare(got, tc.want) != 0 {
+				t.Fatalf("Export() hash = %x, want %x",
+					got, tc.want)
+				return
+			}
+		})
+	}
+
+	t.Run("close without use", func(t *testing.T) {
+		e := seccomp.New(0)
+		if err := e.Close(); !errors.Is(err, syscall.EINVAL) {
+			t.Errorf("Close: error = %v", err)
+			return
+		}
+	})
+
+	t.Run("close partial read", func(t *testing.T) {
+		e := seccomp.New(0)
+		if _, err := e.Read(make([]byte, 0)); err != nil {
+			t.Errorf("Read: error = %v", err)
+			return
+		}
+		if err := e.Close(); err == nil || !errors.Is(err, syscall.ECANCELED) || !errors.Is(err, syscall.EBADF) {
+			t.Errorf("Close: error = %v", err)
+			return
+		}
+	})
+}
+
+func BenchmarkExport(b *testing.B) {
+	buf := make([]byte, 8)
+	for i := 0; i < b.N; i++ {
+		e := seccomp.New(seccomp.FlagExt |
+			seccomp.FlagDenyNS | seccomp.FlagDenyTTY | seccomp.FlagDenyDevel |
+			seccomp.FlagMultiarch | seccomp.FlagLinux32 | seccomp.FlagCan |
+			seccomp.FlagBluetooth)
+		if _, err := io.CopyBuffer(io.Discard, e, buf); err != nil {
+			b.Fatalf("cannot export: %v", err)
+		}
+		if err := e.Close(); err != nil {
+			b.Fatalf("cannot close exporter: %v", err)
+		}
+	}
+}
--- a/sandbox/seccomp/output.go
+++ b/sandbox/seccomp/output.go
@@ -0,0 +1,30 @@
+package seccomp
+
+import "C"
+import "sync/atomic"
+
+var printlnP atomic.Pointer[func(v ...any)]
+
+func SetOutput(f func(v ...any)) {
+	if f == nil {
+		// avoid storing nil function
+		printlnP.Store(nil)
+	} else {
+		printlnP.Store(&f)
+	}
+}
+
+func GetOutput() func(v ...any) {
+	if fp := printlnP.Load(); fp == nil {
+		return nil
+	} else {
+		return *fp
+	}
+}
+
+//export F_println
+func F_println(v *C.char) {
+	if fp := printlnP.Load(); fp != nil {
+		(*fp)(C.GoString(v))
+	}
+}
--- a/sandbox/seccomp/seccomp-build.c
+++ b/sandbox/seccomp/seccomp-build.c
@@ -0,0 +1,302 @@
+#ifndef _GNU_SOURCE
+#define _GNU_SOURCE // CLONE_NEWUSER
+#endif
+
+#include "seccomp-build.h"
+#include <stdlib.h>
+#include <stdio.h>
+#include <assert.h>
+#include <errno.h>
+#include <sys/syscall.h>
+#include <sys/socket.h>
+#include <sys/ioctl.h>
+#include <sys/personality.h>
+#include <sched.h>
+
+#if (SCMP_VER_MAJOR < 2) || \
+    (SCMP_VER_MAJOR == 2 && SCMP_VER_MINOR < 5) || \
+    (SCMP_VER_MAJOR == 2 && SCMP_VER_MINOR == 5 && SCMP_VER_MICRO < 1)
+#error This package requires libseccomp >= v2.5.1
+#endif
+
+struct f_syscall_act {
+  int                  syscall;
+  int                  m_errno;
+  struct scmp_arg_cmp *arg;
+};
+
+#define LEN(arr) (sizeof(arr) / sizeof((arr)[0]))
+
+#define SECCOMP_RULESET_ADD(ruleset) do {                                                                         \
+  if (opts & F_VERBOSE) F_println("adding seccomp ruleset \"" #ruleset "\"");                                     \
+  for (int i = 0; i < LEN(ruleset); i++) {                                                                        \
+    assert(ruleset[i].m_errno == EPERM || ruleset[i].m_errno == ENOSYS);                                          \
+                                                                                                                  \
+    if (ruleset[i].arg)                                                                                           \
+      *ret_p = seccomp_rule_add(ctx, SCMP_ACT_ERRNO(ruleset[i].m_errno), ruleset[i].syscall, 1, *ruleset[i].arg); \
+    else                                                                                                          \
+      *ret_p = seccomp_rule_add(ctx, SCMP_ACT_ERRNO(ruleset[i].m_errno), ruleset[i].syscall, 0);                  \
+                                                                                                                  \
+    if (*ret_p == -EFAULT) {                                                                                      \
+      res = 4;                                                                                                    \
+      goto out;                                                                                                   \
+    } else if (*ret_p < 0) {                                                                                      \
+      res = 5;                                                                                                    \
+      goto out;                                                                                                   \
+    }                                                                                                             \
+  }                                                                                                               \
+} while (0)
+
+int32_t f_build_filter(int *ret_p, int fd, uint32_t arch, uint32_t multiarch, f_syscall_opts opts) {
+  int32_t res = 0; // refer to resErr for meaning
+  int allow_multiarch = opts & F_MULTIARCH;
+  int allowed_personality = PER_LINUX;
+
+  if (opts & F_LINUX32)
+    allowed_personality = PER_LINUX32;
+
+  // flatpak commit 4c3bf179e2e4a2a298cd1db1d045adaf3f564532
+
+  struct f_syscall_act deny_common[] = {
+    // Block dmesg
+    {SCMP_SYS(syslog), EPERM},
+    // Useless old syscall
+    {SCMP_SYS(uselib), EPERM},
+    // Don't allow disabling accounting
+    {SCMP_SYS(acct), EPERM},
+    // Don't allow reading current quota use
+    {SCMP_SYS(quotactl), EPERM},
+
+    // Don't allow access to the kernel keyring
+    {SCMP_SYS(add_key), EPERM},
+    {SCMP_SYS(keyctl), EPERM},
+    {SCMP_SYS(request_key), EPERM},
+
+    // Scary VM/NUMA ops
+    {SCMP_SYS(move_pages), EPERM},
+    {SCMP_SYS(mbind), EPERM},
+    {SCMP_SYS(get_mempolicy), EPERM},
+    {SCMP_SYS(set_mempolicy), EPERM},
+    {SCMP_SYS(migrate_pages), EPERM},
+  };
+
+  // fortify: project-specific extensions
+  struct f_syscall_act deny_common_ext[] = {
+    // system calls for changing the system clock
+    {SCMP_SYS(adjtimex), EPERM},
+    {SCMP_SYS(clock_adjtime), EPERM},
+    {SCMP_SYS(clock_adjtime64), EPERM},
+    {SCMP_SYS(clock_settime), EPERM},
+    {SCMP_SYS(clock_settime64), EPERM},
+    {SCMP_SYS(settimeofday), EPERM},
+
+    // loading and unloading of kernel modules
+    {SCMP_SYS(delete_module), EPERM},
+    {SCMP_SYS(finit_module), EPERM},
+    {SCMP_SYS(init_module), EPERM},
+
+    // system calls for rebooting and reboot preparation
+    {SCMP_SYS(kexec_file_load), EPERM},
+    {SCMP_SYS(kexec_load), EPERM},
+    {SCMP_SYS(reboot), EPERM},
+
+    // system calls for enabling/disabling swap devices
+    {SCMP_SYS(swapoff), EPERM},
+    {SCMP_SYS(swapon), EPERM},
+  };
+
+  struct f_syscall_act deny_ns[] = {
+    // Don't allow subnamespace setups:
+    {SCMP_SYS(unshare), EPERM},
+    {SCMP_SYS(setns), EPERM},
+    {SCMP_SYS(mount), EPERM},
+    {SCMP_SYS(umount), EPERM},
+    {SCMP_SYS(umount2), EPERM},
+    {SCMP_SYS(pivot_root), EPERM},
+    {SCMP_SYS(chroot), EPERM},
+#if defined(__s390__) || defined(__s390x__) || defined(__CRIS__)
+    // Architectures with CONFIG_CLONE_BACKWARDS2: the child stack
+    // and flags arguments are reversed so the flags come second
+    {SCMP_SYS(clone), EPERM, &SCMP_A1(SCMP_CMP_MASKED_EQ, CLONE_NEWUSER, CLONE_NEWUSER)},
+#else
+    // Normally the flags come first
+    {SCMP_SYS(clone), EPERM, &SCMP_A0(SCMP_CMP_MASKED_EQ, CLONE_NEWUSER, CLONE_NEWUSER)},
+#endif
+
+    // seccomp can't look into clone3()'s struct clone_args to check whether
+    // the flags are OK, so we have no choice but to block clone3().
+    // Return ENOSYS so user-space will fall back to clone().
+    // (CVE-2021-41133; see also https://github.com/moby/moby/commit/9f6b562d)
+    {SCMP_SYS(clone3), ENOSYS},
+
+    // New mount manipulation APIs can also change our VFS. There's no
+    // legitimate reason to do these in the sandbox, so block all of them
+    // rather than thinking about which ones might be dangerous.
+    // (CVE-2021-41133)
+    {SCMP_SYS(open_tree), ENOSYS},
+    {SCMP_SYS(move_mount), ENOSYS},
+    {SCMP_SYS(fsopen), ENOSYS},
+    {SCMP_SYS(fsconfig), ENOSYS},
+    {SCMP_SYS(fsmount), ENOSYS},
+    {SCMP_SYS(fspick), ENOSYS},
+    {SCMP_SYS(mount_setattr), ENOSYS},
+  };
+
+  // fortify: project-specific extensions
+  struct f_syscall_act deny_ns_ext[] = {
+    // changing file ownership
+    {SCMP_SYS(chown), EPERM},
+    {SCMP_SYS(chown32), EPERM},
+    {SCMP_SYS(fchown), EPERM},
+    {SCMP_SYS(fchown32), EPERM},
+    {SCMP_SYS(fchownat), EPERM},
+    {SCMP_SYS(lchown), EPERM},
+    {SCMP_SYS(lchown32), EPERM},
+
+    // system calls for changing user ID and group ID credentials
+    {SCMP_SYS(setgid), EPERM},
+    {SCMP_SYS(setgid32), EPERM},
+    {SCMP_SYS(setgroups), EPERM},
+    {SCMP_SYS(setgroups32), EPERM},
+    {SCMP_SYS(setregid), EPERM},
+    {SCMP_SYS(setregid32), EPERM},
+    {SCMP_SYS(setresgid), EPERM},
+    {SCMP_SYS(setresgid32), EPERM},
+    {SCMP_SYS(setresuid), EPERM},
+    {SCMP_SYS(setresuid32), EPERM},
+    {SCMP_SYS(setreuid), EPERM},
+    {SCMP_SYS(setreuid32), EPERM},
+    {SCMP_SYS(setuid), EPERM},
+    {SCMP_SYS(setuid32), EPERM},
+  };
+
+  struct f_syscall_act deny_tty[] = {
+    // Don't allow faking input to the controlling tty (CVE-2017-5226)
+    {SCMP_SYS(ioctl), EPERM, &SCMP_A1(SCMP_CMP_MASKED_EQ, 0xFFFFFFFFu, (int)TIOCSTI)},
+    // In the unlikely event that the controlling tty is a Linux virtual
+    // console (/dev/tty2 or similar), copy/paste operations have an effect
+    // similar to TIOCSTI (CVE-2023-28100)
+    {SCMP_SYS(ioctl), EPERM, &SCMP_A1(SCMP_CMP_MASKED_EQ, 0xFFFFFFFFu, (int)TIOCLINUX)},
+  };
+
+  struct f_syscall_act deny_devel[] = {
+    // Profiling operations; we expect these to be done by tools from outside
+    // the sandbox.  In particular perf has been the source of many CVEs.
+    {SCMP_SYS(perf_event_open), EPERM},
+    // Don't allow you to switch to bsd emulation or whatnot
+    {SCMP_SYS(personality), EPERM, &SCMP_A0(SCMP_CMP_NE, allowed_personality)},
+
+    {SCMP_SYS(ptrace), EPERM}
+  };
+
+  struct f_syscall_act deny_emu[] = {
+    // modify_ldt is a historic source of interesting information leaks,
+    // so it's disabled as a hardening measure.
+    // However, it is required to run old 16-bit applications
+    // as well as some Wine patches, so it's allowed in multiarch.
+    {SCMP_SYS(modify_ldt), EPERM},
+  };
+
+  // fortify: project-specific extensions
+  struct f_syscall_act deny_emu_ext[] = {
+    {SCMP_SYS(subpage_prot), ENOSYS},
+    {SCMP_SYS(switch_endian), ENOSYS},
+    {SCMP_SYS(vm86), ENOSYS},
+    {SCMP_SYS(vm86old), ENOSYS},
+  };
+
+  // Blocklist all but unix, inet, inet6 and netlink
+  struct
+  {
+    int            family;
+    f_syscall_opts flags_mask;
+  } socket_family_allowlist[] = {
+    // NOTE: Keep in numerical order
+    { AF_UNSPEC, 0 },
+    { AF_LOCAL, 0 },
+    { AF_INET, 0 },
+    { AF_INET6, 0 },
+    { AF_NETLINK, 0 },
+    { AF_CAN, F_CAN },
+    { AF_BLUETOOTH, F_BLUETOOTH },
+  };
+
+  scmp_filter_ctx ctx = seccomp_init(SCMP_ACT_ALLOW);
+  if (ctx == NULL) {
+    res = 1;
+    goto out;
+  } else
+    errno = 0;
+
+  // We only really need to handle arches on multiarch systems.
+  // If only one arch is supported the default is fine
+  if (arch != 0) {
+    // This *adds* the target arch, instead of replacing the
+    // native one. This is not ideal, because we'd like to only
+    // allow the target arch, but we can't really disallow the
+    // native arch at this point, because then bubblewrap
+    // couldn't continue running.
+    *ret_p = seccomp_arch_add(ctx, arch);
+    if (*ret_p < 0 && *ret_p != -EEXIST) {
+      res = 2;
+      goto out;
+    }
+
+    if (allow_multiarch && multiarch != 0) {
+      *ret_p = seccomp_arch_add(ctx, multiarch);
+      if (*ret_p < 0 && *ret_p != -EEXIST) {
+        res = 3;
+        goto out;
+      }
+    }
+  }
+
+  SECCOMP_RULESET_ADD(deny_common);
+  if (opts & F_DENY_NS) SECCOMP_RULESET_ADD(deny_ns);
+  if (opts & F_DENY_TTY) SECCOMP_RULESET_ADD(deny_tty);
+  if (opts & F_DENY_DEVEL) SECCOMP_RULESET_ADD(deny_devel);
+  if (!allow_multiarch) SECCOMP_RULESET_ADD(deny_emu);
+  if (opts & F_EXT) {
+    SECCOMP_RULESET_ADD(deny_common_ext);
+    if (opts & F_DENY_NS) SECCOMP_RULESET_ADD(deny_ns_ext);
+    if (!allow_multiarch) SECCOMP_RULESET_ADD(deny_emu_ext);
+  }
+
+  // Socket filtering doesn't work on e.g. i386, so ignore failures here
+  // However, we need to user seccomp_rule_add_exact to avoid libseccomp doing
+  // something else: https://github.com/seccomp/libseccomp/issues/8
+  int last_allowed_family = -1;
+  for (int i = 0; i < LEN(socket_family_allowlist); i++) {
+    if (socket_family_allowlist[i].flags_mask != 0 &&
+        (socket_family_allowlist[i].flags_mask & opts) != socket_family_allowlist[i].flags_mask)
+      continue;
+
+    for (int disallowed = last_allowed_family + 1; disallowed < socket_family_allowlist[i].family; disallowed++) {
+      // Blocklist the in-between valid families
+      seccomp_rule_add_exact(ctx, SCMP_ACT_ERRNO(EAFNOSUPPORT), SCMP_SYS(socket), 1, SCMP_A0(SCMP_CMP_EQ, disallowed));
+    }
+    last_allowed_family = socket_family_allowlist[i].family;
+  }
+  // Blocklist the rest
+  seccomp_rule_add_exact(ctx, SCMP_ACT_ERRNO(EAFNOSUPPORT), SCMP_SYS(socket), 1, SCMP_A0(SCMP_CMP_GE, last_allowed_family + 1));
+
+  if (fd < 0) {
+    *ret_p = seccomp_load(ctx);
+    if (*ret_p != 0) {
+      res = 7;
+      goto out;
+    }
+  } else {
+    *ret_p = seccomp_export_bpf(ctx, fd);
+    if (*ret_p != 0) {
+      res = 6;
+      goto out;
+    }
+  }
+
+out:
+  if (ctx)
+    seccomp_release(ctx);
+
+  return res;
+}
--- a/sandbox/seccomp/seccomp-build.h
+++ b/sandbox/seccomp/seccomp-build.h
@@ -0,0 +1,23 @@
+#include <stdint.h>
+#include <seccomp.h>
+
+#if (SCMP_VER_MAJOR < 2) || \
+    (SCMP_VER_MAJOR == 2 && SCMP_VER_MINOR < 5) || \
+    (SCMP_VER_MAJOR == 2 && SCMP_VER_MINOR == 5 && SCMP_VER_MICRO < 1)
+#error This package requires libseccomp >= v2.5.1
+#endif
+
+typedef enum {
+  F_VERBOSE    = 1 << 0,
+  F_EXT        = 1 << 1,
+  F_DENY_NS    = 1 << 2,
+  F_DENY_TTY   = 1 << 3,
+  F_DENY_DEVEL = 1 << 4,
+  F_MULTIARCH  = 1 << 5,
+  F_LINUX32    = 1 << 6,
+  F_CAN        = 1 << 7,
+  F_BLUETOOTH  = 1 << 8,
+} f_syscall_opts;
+
+extern void F_println(char *v);
+int32_t f_build_filter(int *ret_p, int fd, uint32_t arch, uint32_t multiarch, f_syscall_opts opts);
--- a/sandbox/seccomp/seccomp.go
+++ b/sandbox/seccomp/seccomp.go
@@ -0,0 +1,114 @@
+package seccomp
+
+/*
+#cgo linux pkg-config: --static libseccomp
+
+#include "seccomp-build.h"
+*/
+import "C"
+import (
+	"errors"
+	"fmt"
+	"runtime"
+	"syscall"
+)
+
+// LibraryError represents a libseccomp error.
+type LibraryError struct {
+	Prefix  string
+	Seccomp syscall.Errno
+	Errno   error
+}
+
+func (e *LibraryError) Error() string {
+	if e.Seccomp == 0 {
+		if e.Errno == nil {
+			panic("invalid libseccomp error")
+		}
+		return fmt.Sprintf("%s: %s", e.Prefix, e.Errno)
+	}
+	if e.Errno == nil {
+		return fmt.Sprintf("%s: %s", e.Prefix, e.Seccomp)
+	}
+	return fmt.Sprintf("%s: %s (%s)", e.Prefix, e.Seccomp, e.Errno)
+}
+
+func (e *LibraryError) Is(err error) bool {
+	if e == nil {
+		return err == nil
+	}
+	if ef, ok := err.(*LibraryError); ok {
+		return *e == *ef
+	}
+	return (e.Seccomp != 0 && errors.Is(err, e.Seccomp)) ||
+		(e.Errno != nil && errors.Is(err, e.Errno))
+}
+
+var resPrefix = [...]string{
+	0: "",
+	1: "seccomp_init failed",
+	2: "seccomp_arch_add failed",
+	3: "seccomp_arch_add failed (multiarch)",
+	4: "internal libseccomp failure",
+	5: "seccomp_rule_add failed",
+	6: "seccomp_export_bpf failed",
+	7: "seccomp_load failed",
+}
+
+type SyscallOpts = C.f_syscall_opts
+
+const (
+	flagVerbose SyscallOpts = C.F_VERBOSE
+	// FlagExt are project-specific extensions.
+	FlagExt SyscallOpts = C.F_EXT
+	// FlagDenyNS denies namespace setup syscalls.
+	FlagDenyNS SyscallOpts = C.F_DENY_NS
+	// FlagDenyTTY denies faking input.
+	FlagDenyTTY SyscallOpts = C.F_DENY_TTY
+	// FlagDenyDevel denies development-related syscalls.
+	FlagDenyDevel SyscallOpts = C.F_DENY_DEVEL
+	// FlagMultiarch allows multiarch/emulation.
+	FlagMultiarch SyscallOpts = C.F_MULTIARCH
+	// FlagLinux32 sets PER_LINUX32.
+	FlagLinux32 SyscallOpts = C.F_LINUX32
+	// FlagCan allows AF_CAN.
+	FlagCan SyscallOpts = C.F_CAN
+	// FlagBluetooth allows AF_BLUETOOTH.
+	FlagBluetooth SyscallOpts = C.F_BLUETOOTH
+)
+
+func buildFilter(fd int, opts SyscallOpts) error {
+	var (
+		arch      C.uint32_t = 0
+		multiarch C.uint32_t = 0
+	)
+	switch runtime.GOARCH {
+	case "386":
+		arch = C.SCMP_ARCH_X86
+	case "amd64":
+		arch = C.SCMP_ARCH_X86_64
+		multiarch = C.SCMP_ARCH_X86
+	case "arm":
+		arch = C.SCMP_ARCH_ARM
+	case "arm64":
+		arch = C.SCMP_ARCH_AARCH64
+		multiarch = C.SCMP_ARCH_ARM
+	}
+
+	// this removes repeated transitions between C and Go execution
+	// when producing log output via F_println and CPrintln is nil
+	if fp := printlnP.Load(); fp != nil {
+		opts |= flagVerbose
+	}
+
+	var ret C.int
+	res, err := C.f_build_filter(&ret, C.int(fd), arch, multiarch, opts)
+	if prefix := resPrefix[res]; prefix != "" {
+		return &LibraryError{
+			prefix,
+			-syscall.Errno(ret),
+			err,
+		}
+	}
+	return err
+}
--- a/sandbox/seccomp/seccomp_test.go
+++ b/sandbox/seccomp/seccomp_test.go
@@ -0,0 +1,65 @@
+package seccomp_test
+
+import (
+	"errors"
+	"runtime"
+	"syscall"
+	"testing"
+
+	"git.gensokyo.uk/security/fortify/sandbox/seccomp"
+)
+
+func TestLibraryError(t *testing.T) {
+	testCases := []struct {
+		name    string
+		sample  *seccomp.LibraryError
+		want    string
+		wantIs  bool
+		compare error
+	}{
+		{
+			"full",
+			&seccomp.LibraryError{Prefix: "seccomp_export_bpf failed", Seccomp: syscall.ECANCELED, Errno: syscall.EBADF},
+			"seccomp_export_bpf failed: operation canceled (bad file descriptor)",
+			true,
+			&seccomp.LibraryError{Prefix: "seccomp_export_bpf failed", Seccomp: syscall.ECANCELED, Errno: syscall.EBADF},
+		},
+		{
+			"errno only",
+			&seccomp.LibraryError{Prefix: "seccomp_init failed", Errno: syscall.ENOMEM},
+			"seccomp_init failed: cannot allocate memory",
+			false,
+			nil,
+		},
+		{
+			"seccomp only",
+			&seccomp.LibraryError{Prefix: "internal libseccomp failure", Seccomp: syscall.EFAULT},
+			"internal libseccomp failure: bad address",
+			true,
+			syscall.EFAULT,
+		},
+	}
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			if errors.Is(tc.sample, tc.compare) != tc.wantIs {
+				t.Errorf("errors.Is(%#v, %#v) did not return %v",
+					tc.sample, tc.compare, tc.wantIs)
+			}
+
+			if got := tc.sample.Error(); got != tc.want {
+				t.Errorf("Error: %q, want %q",
+					got, tc.want)
+			}
+		})
+	}
+
+	t.Run("invalid", func(t *testing.T) {
+		wantPanic := "invalid libseccomp error"
+		defer func() {
+			if r := recover(); r != wantPanic {
+				t.Errorf("panic: %q, want %q", r, wantPanic)
+			}
+		}()
+		runtime.KeepAlive(new(seccomp.LibraryError).Error())
+	})
+}
--- a/sandbox/sequential.go
+++ b/sandbox/sequential.go
@@ -0,0 +1,180 @@
+package sandbox
+
+import (
+	"encoding/gob"
+	"fmt"
+	"math"
+	"os"
+	"path"
+	"syscall"
+	"unsafe"
+)
+
+func init() { gob.Register(new(BindMount)) }
+
+// BindMount bind mounts host path Source on container path Target.
+type BindMount struct {
+	Source, Target string
+
+	Flags int
+}
+
+func (b *BindMount) apply(*InitParams) error {
+	if !path.IsAbs(b.Source) || !path.IsAbs(b.Target) {
+		return msg.WrapErr(syscall.EBADE,
+			"path is not absolute")
+	}
+	return bindMount(b.Source, b.Target, b.Flags)
+}
+
+func (b *BindMount) Is(op Op) bool { vb, ok := op.(*BindMount); return ok && *b == *vb }
+func (b *BindMount) String() string {
+	if b.Source == b.Target {
+		return fmt.Sprintf("%q flags %#x", b.Source, b.Flags)
+	}
+	return fmt.Sprintf("%q on %q flags %#x", b.Source, b.Target, b.Flags&BindWritable)
+}
+func (f *Ops) Bind(source, target string, flags int) *Ops {
+	*f = append(*f, &BindMount{source, target, flags | BindRecursive})
+	return f
+}
+
+func init() { gob.Register(new(MountProc)) }
+
+// MountProc mounts a private proc instance on container Path.
+type MountProc struct {
+	Path string
+}
+
+func (p *MountProc) apply(*InitParams) error {
+	if !path.IsAbs(p.Path) {
+		return msg.WrapErr(syscall.EBADE,
+			fmt.Sprintf("path %q is not absolute", p.Path))
+	}
+
+	target := toSysroot(p.Path)
+	if err := os.MkdirAll(target, 0755); err != nil {
+		return msg.WrapErr(err, err.Error())
+	}
+	return wrapErrSuffix(syscall.Mount("proc", target, "proc",
+		syscall.MS_NOSUID|syscall.MS_NOEXEC|syscall.MS_NODEV, ""),
+		fmt.Sprintf("cannot mount proc on %q:", p.Path))
+}
+
+func init() { gob.Register(new(MountDev)) }
+
+// MountDev mounts dev on container Path.
+type MountDev struct {
+	Path string
+}
+
+func (d *MountDev) apply(params *InitParams) error {
+	if !path.IsAbs(d.Path) {
+		return msg.WrapErr(syscall.EBADE,
+			fmt.Sprintf("path %q is not absolute", d.Path))
+	}
+	target := toSysroot(d.Path)
+
+	if err := mountTmpfs("devtmpfs", d.Path, 0, 0755); err != nil {
+		return err
+	}
+
+	for _, name := range []string{"null", "zero", "full", "random", "urandom", "tty"} {
+		if err := bindMount(
+			"/dev/"+name, path.Join(d.Path, name),
+			BindSource|BindDevices,
+		); err != nil {
+			return err
+		}
+	}
+	for i, name := range []string{"stdin", "stdout", "stderr"} {
+		if err := os.Symlink(
+			"/proc/self/fd/"+string(rune(i+'0')),
+			path.Join(target, name),
+		); err != nil {
+			return msg.WrapErr(err, err.Error())
+		}
+	}
+	for _, pair := range [][2]string{
+		{"/proc/self/fd", "fd"},
+		{"/proc/kcore", "core"},
+		{"pts/ptmx", "ptmx"},
+	} {
+		if err := os.Symlink(pair[0], path.Join(target, pair[1])); err != nil {
+			return msg.WrapErr(err, err.Error())
+		}
+	}
+
+	devPtsPath := path.Join(target, "pts")
+	for _, name := range []string{path.Join(target, "shm"), devPtsPath} {
+		if err := os.Mkdir(name, 0755); err != nil {
+			return msg.WrapErr(err, err.Error())
+		}
+	}
+
+	if err := syscall.Mount("devpts", devPtsPath, "devpts",
+		syscall.MS_NOSUID|syscall.MS_NOEXEC,
+		"newinstance,ptmxmode=0666,mode=620"); err != nil {
+		return wrapErrSuffix(err,
+			fmt.Sprintf("cannot mount devpts on %q:", devPtsPath))
+	}
+
+	if params.Flags&FAllowTTY != 0 {
+		var buf [8]byte
+		if _, _, errno := syscall.Syscall(
+			syscall.SYS_IOCTL, 1, syscall.TIOCGWINSZ,
+			uintptr(unsafe.Pointer(&buf[0])),
+		); errno == 0 {
+			if err := bindMount(
+				"/proc/self/fd/1", path.Join(d.Path, "console"),
+				BindDevices,
+			); err != nil {
+				return err
+			}
+		}
+	}
+
+	return nil
+}
+
+func (d *MountDev) Is(op Op) bool  { vd, ok := op.(*MountDev); return ok && *d == *vd }
+func (d *MountDev) String() string { return fmt.Sprintf("dev on %q", d.Path) }
+func (f *Ops) Dev(dest string) *Ops {
+	*f = append(*f, &MountDev{dest})
+	return f
+}
+
+func (p *MountProc) Is(op Op) bool  { vp, ok := op.(*MountProc); return ok && *p == *vp }
+func (p *MountProc) String() string { return fmt.Sprintf("proc on %q", p.Path) }
+func (f *Ops) Proc(dest string) *Ops {
+	*f = append(*f, &MountProc{dest})
+	return f
+}
+
+func init() { gob.Register(new(MountTmpfs)) }
+
+// MountTmpfs mounts tmpfs on container Path.
+type MountTmpfs struct {
+	Path string
+	Size int
+	Perm os.FileMode
+}
+
+func (t *MountTmpfs) apply(*InitParams) error {
+	if !path.IsAbs(t.Path) {
+		return msg.WrapErr(syscall.EBADE,
+			fmt.Sprintf("path %q is not absolute", t.Path))
+	}
+	if t.Size < 0 || t.Size > math.MaxUint>>1 {
+		return msg.WrapErr(syscall.EBADE,
+			fmt.Sprintf("size %d out of bounds", t.Size))
+	}
+	return mountTmpfs("tmpfs", t.Path, t.Size, t.Perm)
+}
+
+func (t *MountTmpfs) Is(op Op) bool  { vt, ok := op.(*MountTmpfs); return ok && *t == *vt }
+func (t *MountTmpfs) String() string { return fmt.Sprintf("tmpfs on %q size %d", t.Path, t.Size) }
+func (f *Ops) Tmpfs(dest string, size int, perm os.FileMode) *Ops {
+	*f = append(*f, &MountTmpfs{dest, size, perm})
+	return f
+}
--- a/sandbox/syscall.go
+++ b/sandbox/syscall.go
@@ -0,0 +1,41 @@
+package sandbox
+
+import "syscall"
+
+const (
+	SUID_DUMP_DISABLE = iota
+	SUID_DUMP_USER
+)
+
+func SetDumpable(dumpable uintptr) error {
+	// linux/sched/coredump.h
+	if _, _, errno := syscall.RawSyscall(syscall.SYS_PRCTL, syscall.PR_SET_DUMPABLE, dumpable, 0); errno != 0 {
+		return errno
+	}
+
+	return nil
+}
+
+func SetPdeathsig(sig syscall.Signal) error {
+	if _, _, errno := syscall.RawSyscall(syscall.SYS_PRCTL, syscall.PR_SET_PDEATHSIG, uintptr(sig), 0); errno != 0 {
+		return errno
+	}
+
+	return nil
+}
+
+// IgnoringEINTR makes a function call and repeats it if it returns an
+// EINTR error. This appears to be required even though we install all
+// signal handlers with SA_RESTART: see #22838, #38033, #38836, #40846.
+// Also #20400 and #36644 are issues in which a signal handler is
+// installed without setting SA_RESTART. None of these are the common case,
+// but there are enough of them that it seems that we can't avoid
+// an EINTR loop.
+func IgnoringEINTR(fn func() error) error {
+	for {
+		err := fn()
+		if err != syscall.EINTR {
+			return err
+		}
+	}
+}