seccomp: move out of helper

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

7 files changed, 781 insertions, 0 deletions

diff --git a/seccomp/api.go b/seccomp/api.go
new file mode 100644
index 00000000..697b09f3
--- /dev/null
+++ b/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
+}
diff --git a/seccomp/export.go b/seccomp/export.go
new file mode 100644
index 00000000..95684235
--- /dev/null
+++ b/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}
+}
diff --git a/seccomp/export_test.go b/seccomp/export_test.go
new file mode 100644
index 00000000..991517f4
--- /dev/null
+++ b/seccomp/export_test.go
@@ -0,0 +1,139 @@
+package seccomp_test
+
+import (
+	"crypto/sha512"
+	"errors"
+	"io"
+	"log"
+	"slices"
+	"syscall"
+	"testing"
+
+	"git.gensokyo.uk/security/fortify/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) {
+			seccomp.CPrintln = log.Println
+			t.Cleanup(func() { seccomp.CPrintln = nil })
+
+			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)
+		}
+	}
+}
diff --git a/seccomp/seccomp-build.c b/seccomp/seccomp-build.c
new file mode 100644
index 00000000..4b2114f4
--- /dev/null
+++ b/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;
+}
diff --git a/seccomp/seccomp-build.h b/seccomp/seccomp-build.h
new file mode 100644
index 00000000..7f1be761
--- /dev/null
+++ b/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);
\ No newline at end of file
diff --git a/seccomp/seccomp.go b/seccomp/seccomp.go
new file mode 100644
index 00000000..05c314c9
--- /dev/null
+++ b/seccomp/seccomp.go
@@ -0,0 +1,123 @@
+package seccomp
+
+/*
+#cgo linux pkg-config: --static libseccomp
+
+#include "seccomp-build.h"
+*/
+import "C"
+import (
+	"errors"
+	"fmt"
+	"runtime"
+	"syscall"
+)
+
+var CPrintln func(v ...any)
+
+// 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 CPrintln != 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
+}
+
+//export F_println
+func F_println(v *C.char) {
+	if CPrintln != nil {
+		CPrintln(C.GoString(v))
+	}
+}
diff --git a/seccomp/seccomp_test.go b/seccomp/seccomp_test.go
new file mode 100644
index 00000000..f2949ef8
--- /dev/null
+++ b/seccomp/seccomp_test.go
@@ -0,0 +1,65 @@
+package seccomp_test
+
+import (
+	"errors"
+	"runtime"
+	"syscall"
+	"testing"
+
+	"git.gensokyo.uk/security/fortify/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())
+	})
+}