internal/pkg: implement caching for files

This change contains primitives for validating and caching single-file artifacts.

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

.clang-format

@@ -0,0 +1,2 @@
+ColumnLimit: 0
+IndentWidth: 4

.gitea/workflows/test.yml

@@ -2,7 +2,6 @@ name: Test
 
 on:
   - push
-  - pull_request
 
 jobs:
   hakurei:
@@ -73,6 +72,23 @@ jobs:
           path: result/*
           retention-days: 1
 
+  sharefs:
+    name: ShareFS
+    runs-on: nix
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Run NixOS test
+        run: nix build --out-link "result" --print-out-paths --print-build-logs .#checks.x86_64-linux.sharefs
+
+      - name: Upload test output
+        uses: actions/upload-artifact@v3
+        with:
+          name: "sharefs-vm-output"
+          path: result/*
+          retention-days: 1
+
   hpkg:
     name: Hpkg
     runs-on: nix
@@ -97,6 +113,7 @@ jobs:
       - race
       - sandbox
       - sandbox-race
+      - sharefs
       - hpkg
     runs-on: nix
     steps:

cmd/hakurei/command.go

@@ -11,21 +11,24 @@ import (
 	"strconv"
 	"sync"
 	"time"
-	_ "unsafe"
+	_ "unsafe" // for go:linkname
 
 	"hakurei.app/command"
 	"hakurei.app/container/check"
 	"hakurei.app/container/fhs"
 	"hakurei.app/hst"
-	"hakurei.app/internal"
+	"hakurei.app/internal/dbus"
 	"hakurei.app/internal/env"
+	"hakurei.app/internal/info"
 	"hakurei.app/internal/outcome"
 	"hakurei.app/message"
-	"hakurei.app/system/dbus"
 )
 
+// optionalErrorUnwrap calls [errors.Unwrap] and returns the resulting value
+// if it is not nil, or the original value if it is.
+//
 //go:linkname optionalErrorUnwrap hakurei.app/container.optionalErrorUnwrap
-func optionalErrorUnwrap(_ error) error
+func optionalErrorUnwrap(err error) error
 
 func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErrs, out io.Writer) command.Command {
 	var (
@@ -88,7 +91,7 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 
 			flagPrivateRuntime, flagPrivateTmpdir bool
 
-			flagWayland, flagX11, flagDBus, flagPulse bool
+			flagWayland, flagX11, flagDBus, flagPipeWire, flagPulse bool
 		)
 
 		c.NewCommand("run", "Configure and start a permissive container", func(args []string) error {
@@ -143,8 +146,8 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 			if flagDBus {
 				et |= hst.EDBus
 			}
-			if flagPulse {
-				et |= hst.EPulse
+			if flagPipeWire || flagPulse {
+				et |= hst.EPipeWire
 			}
 
 			config := &hst.Config{
@@ -294,8 +297,10 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 				"Enable direct connection to X11").
 			Flag(&flagDBus, "dbus", command.BoolFlag(false),
 				"Enable proxied connection to D-Bus").
+			Flag(&flagPipeWire, "pipewire", command.BoolFlag(false),
+				"Enable connection to PipeWire via SecurityContext").
 			Flag(&flagPulse, "pulse", command.BoolFlag(false),
-				"Enable direct connection to PulseAudio")
+				"Enable PulseAudio compatibility daemon")
 	}
 
 	{
@@ -350,7 +355,7 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 		}).Flag(&flagShort, "short", command.BoolFlag(false), "Print instance id")
 	}
 
-	c.Command("version", "Display version information", func(args []string) error { fmt.Println(internal.Version()); return errSuccess })
+	c.Command("version", "Display version information", func(args []string) error { fmt.Println(info.Version()); return errSuccess })
 	c.Command("license", "Show full license text", func(args []string) error { fmt.Println(license); return errSuccess })
 	c.Command("template", "Produce a config template", func(args []string) error { encodeJSON(log.Fatal, os.Stdout, false, hst.Template()); return errSuccess })
 	c.Command("help", "Show this help message", func([]string) error { c.PrintHelp(); return errSuccess })

cmd/hakurei/command_test.go

@@ -36,7 +36,7 @@ Commands:
 		},
 		{
 			"run", []string{"run", "-h"}, `
-Usage:	hakurei run [-h | --help] [--dbus-config <value>] [--dbus-system <value>] [--mpris] [--dbus-log] [--id <value>] [-a <int>] [-g <value>] [-d <value>] [-u <value>] [--private-runtime] [--private-tmpdir] [--wayland] [-X] [--dbus] [--pulse] COMMAND [OPTIONS]
+Usage:	hakurei run [-h | --help] [--dbus-config <value>] [--dbus-system <value>] [--mpris] [--dbus-log] [--id <value>] [-a <int>] [-g <value>] [-d <value>] [-u <value>] [--private-runtime] [--private-tmpdir] [--wayland] [-X] [--dbus] [--pipewire] [--pulse] COMMAND [OPTIONS]
 
 Flags:
   -X	Enable direct connection to X11
@@ -58,12 +58,14 @@ Flags:
     	Reverse-DNS style Application identifier, leave empty to inherit instance identifier
   -mpris
     	Allow owning MPRIS D-Bus path, has no effect if custom config is available
+  -pipewire
+    	Enable connection to PipeWire via SecurityContext
   -private-runtime
     	Do not share XDG_RUNTIME_DIR between containers under the same identity
   -private-tmpdir
     	Do not share TMPDIR between containers under the same identity
   -pulse
-    	Enable direct connection to PulseAudio
+    	Enable PulseAudio compatibility daemon
   -u string
     	Passwd user name within sandbox (default "chronos")
   -wayland

cmd/hakurei/json_test.go

@@ -1,18 +1,13 @@
-package main_test
+package main
 
 import (
-	"io"
 	"reflect"
 	"strings"
 	"testing"
-	_ "unsafe"
 
 	"hakurei.app/container/stub"
 )
 
-//go:linkname decodeJSON hakurei.app/cmd/hakurei.decodeJSON
-func decodeJSON(fatal func(v ...any), op string, r io.Reader, v any)
-
 func TestDecodeJSON(t *testing.T) {
 	t.Parallel()
 
@@ -62,9 +57,6 @@ func TestDecodeJSON(t *testing.T) {
 	}
 }
 
-//go:linkname encodeJSON hakurei.app/cmd/hakurei.encodeJSON
-func encodeJSON(fatal func(v ...any), output io.Writer, short bool, v any)
-
 func TestEncodeJSON(t *testing.T) {
 	t.Parallel()
 
@@ -74,7 +66,7 @@ func TestEncodeJSON(t *testing.T) {
 		want string
 	}{
 		{"marshaler", errorJSONMarshaler{},
-			`cannot encode json for main_test.errorJSONMarshaler: unique error 3735928559 injected by the test suite`},
+			`cannot encode json for main.errorJSONMarshaler: unique error 3735928559 injected by the test suite`},
 		{"default", func() {},
 			`cannot write json: json: unsupported type: func()`},
 	}

cmd/hakurei/print.go

@@ -12,8 +12,6 @@ import (
 	"time"
 
 	"hakurei.app/hst"
-	"hakurei.app/internal"
-	"hakurei.app/internal/env"
 	"hakurei.app/internal/outcome"
 	"hakurei.app/internal/store"
 	"hakurei.app/message"
@@ -23,21 +21,19 @@ import (
 func printShowSystem(output io.Writer, short, flagJSON bool) {
 	t := newPrinter(output)
 	defer t.MustFlush()
-
-	info := &hst.Info{Version: internal.Version(), User: new(outcome.Hsu).MustID(nil)}
-	env.CopyPaths().Copy(&info.Paths, info.User)
+	hi := outcome.Info()
 
 	if flagJSON {
-		encodeJSON(log.Fatal, output, short, info)
+		encodeJSON(log.Fatal, output, short, hi)
 		return
 	}
 
-	t.Printf("Version:\t%s\n", info.Version)
-	t.Printf("User:\t%d\n", info.User)
-	t.Printf("TempDir:\t%s\n", info.TempDir)
-	t.Printf("SharePath:\t%s\n", info.SharePath)
-	t.Printf("RuntimePath:\t%s\n", info.RuntimePath)
-	t.Printf("RunDirPath:\t%s\n", info.RunDirPath)
+	t.Printf("Version:\t%s (libwayland %s)\n", hi.Version, hi.WaylandVersion)
+	t.Printf("User:\t%d\n", hi.User)
+	t.Printf("TempDir:\t%s\n", hi.TempDir)
+	t.Printf("SharePath:\t%s\n", hi.SharePath)
+	t.Printf("RuntimePath:\t%s\n", hi.RuntimePath)
+	t.Printf("RunDirPath:\t%s\n", hi.RunDirPath)
 }
 
 // printShowInstance writes a representation of [hst.State] or [hst.Config] to output.
@@ -90,12 +86,6 @@ func printShowInstance(
 		t.Printf(" Groups:\t%s\n", strings.Join(config.Groups, ", "))
 	}
 	if config.Container != nil {
-		if config.Container.Home != nil {
-			t.Printf(" Home:\t%s\n", config.Container.Home)
-		}
-		if config.Container.Hostname != "" {
-			t.Printf(" Hostname:\t%s\n", config.Container.Hostname)
-		}
 		flags := config.Container.Flags.String()
 
 		// this is included in the upper hst.Config struct but is relevant here
@@ -110,6 +100,12 @@ func printShowInstance(
 		}
 		t.Printf(" Flags:\t%s\n", flags)
 
+		if config.Container.Home != nil {
+			t.Printf(" Home:\t%s\n", config.Container.Home)
+		}
+		if config.Container.Hostname != "" {
+			t.Printf(" Hostname:\t%s\n", config.Container.Hostname)
+		}
 		if config.Container.Path != nil {
 			t.Printf(" Path:\t%s\n", config.Container.Path)
 		}

cmd/hakurei/print_test.go

@@ -32,7 +32,7 @@ var (
 		PID:     0xbeef,
 		ShimPID: 0xcafe,
 		Config: &hst.Config{
-			Enablements: hst.NewEnablements(hst.EWayland | hst.EPulse),
+			Enablements: hst.NewEnablements(hst.EWayland | hst.EPipeWire),
 			Identity:    1,
 			Container: &hst.ContainerConfig{
 				Shell: check.MustAbs("/bin/sh"),
@@ -62,11 +62,11 @@ func TestPrintShowInstance(t *testing.T) {
 		{"nil", nil, nil, false, false, "Error: invalid configuration!\n\n", false},
 		{"config", nil, hst.Template(), false, false, `App
  Identity:       9 (org.chromium.Chromium)
- Enablements:    wayland, dbus, pulseaudio
+ Enablements:    wayland, dbus, pipewire
  Groups:         video, dialout, plugdev
+ Flags:          multiarch, compat, devel, userns, net, abstract, tty, mapuid, device, runtime, tmpdir
  Home:           /data/data/org.chromium.Chromium
  Hostname:       localhost
- Flags:          multiarch, compat, devel, userns, net, abstract, tty, mapuid, device, runtime, tmpdir
  Path:           /run/current-system/sw/bin/chromium
  Arguments:      chromium --ignore-gpu-blocklist --disable-smooth-scrolling --enable-features=UseOzonePlatform --ozone-platform=wayland
 
@@ -159,11 +159,11 @@ Session bus
 
 App
  Identity:       9 (org.chromium.Chromium)
- Enablements:    wayland, dbus, pulseaudio
+ Enablements:    wayland, dbus, pipewire
  Groups:         video, dialout, plugdev
+ Flags:          multiarch, compat, devel, userns, net, abstract, tty, mapuid, device, runtime, tmpdir
  Home:           /data/data/org.chromium.Chromium
  Hostname:       localhost
- Flags:          multiarch, compat, devel, userns, net, abstract, tty, mapuid, device, runtime, tmpdir
  Path:           /run/current-system/sw/bin/chromium
  Arguments:      chromium --ignore-gpu-blocklist --disable-smooth-scrolling --enable-features=UseOzonePlatform --ozone-platform=wayland
 
@@ -215,7 +215,7 @@ App
   "enablements": {
     "wayland": true,
     "dbus": true,
-    "pulse": true
+    "pipewire": true
   },
   "session_bus": {
     "see": null,
@@ -366,7 +366,7 @@ App
   "enablements": {
     "wayland": true,
     "dbus": true,
-    "pulse": true
+    "pipewire": true
   },
   "session_bus": {
     "see": null,
@@ -564,7 +564,7 @@ func TestPrintPs(t *testing.T) {
     "enablements": {
       "wayland": true,
       "dbus": true,
-      "pulse": true
+      "pipewire": true
     },
     "session_bus": {
       "see": null,
@@ -715,7 +715,7 @@ func TestPrintPs(t *testing.T) {
     "shim_pid": 51966,
     "enablements": {
       "wayland": true,
-      "pulse": true
+      "pipewire": true
     },
     "identity": 1,
     "groups": null,

cmd/hpkg/build.nix

@@ -45,7 +45,7 @@
   allow_wayland ? true,
   allow_x11 ? false,
   allow_dbus ? true,
-  allow_pulse ? true,
+  allow_audio ? true,
   gpu ? allow_wayland || allow_x11,
 }:
 
@@ -175,7 +175,7 @@ let
       wayland = allow_wayland;
       x11 = allow_x11;
       dbus = allow_dbus;
-      pulse = allow_pulse;
+      pipewire = allow_audio;
     };
 
     mesa = if gpu then mesaWrappers else null;

cmd/hpkg/proc.go

@@ -10,11 +10,11 @@ import (
 	"os/exec"
 
 	"hakurei.app/hst"
-	"hakurei.app/internal"
+	"hakurei.app/internal/info"
 	"hakurei.app/message"
 )
 
-var hakureiPathVal = internal.MustHakureiPath().String()
+var hakureiPathVal = info.MustHakureiPath().String()
 
 func mustRunApp(ctx context.Context, msg message.Msg, config *hst.Config, beforeFail func()) {
 	var (

cmd/hpkg/test/default.nix

@@ -1,5 +1,5 @@
 {
-  nixosTest,
+  testers,
   callPackage,
 
   system,
@@ -8,7 +8,7 @@
 let
   buildPackage = self.buildPackage.${system};
 in
-nixosTest {
+testers.nixosTest {
   name = "hpkg";
   nodes.machine = {
     environment.etc = {

cmd/hpkg/test/test.py

@@ -90,13 +90,13 @@ wait_for_window("hakurei@machine-foot")
 machine.send_chars("clear; wayland-info && touch /tmp/success-client\n")
 machine.wait_for_file("/tmp/hakurei.0/tmpdir/2/success-client")
 collect_state_ui("app_wayland")
-check_state("foot", {"wayland": True, "dbus": True, "pulse": True})
+check_state("foot", {"wayland": True, "dbus": True, "pipewire": True})
 # Verify acl on XDG_RUNTIME_DIR:
-print(machine.succeed("getfacl --absolute-names --omit-header --numeric /run/user/1000 | grep 10002"))
+print(machine.succeed("getfacl --absolute-names --omit-header --numeric /tmp/hakurei.0/runtime | grep 10002"))
 machine.send_chars("exit\n")
 machine.wait_until_fails("pgrep foot")
 # Verify acl cleanup on XDG_RUNTIME_DIR:
-machine.wait_until_fails("getfacl --absolute-names --omit-header --numeric /run/user/1000 | grep 10002")
+machine.wait_until_fails("getfacl --absolute-names --omit-header --numeric /tmp/hakurei.0/runtime | grep 10002")
 
 # Exit Sway and verify process exit status 0:
 swaymsg("exit", succeed=False)
@@ -107,4 +107,4 @@ print(machine.succeed("find /tmp/hakurei.0 "
     + "-path '/tmp/hakurei.0/runtime/*/*' -prune -o "
     + "-path '/tmp/hakurei.0/tmpdir/*/*' -prune -o "
     + "-print"))
-print(machine.succeed("find /run/user/1000/hakurei"))
+print(machine.fail("ls /run/user/1000/hakurei"))

cmd/sharefs/fuse-operations.c

@@ -0,0 +1,282 @@
+#ifndef _GNU_SOURCE
+#define _GNU_SOURCE /* O_DIRECT */
+#endif
+
+#include <dirent.h>
+#include <errno.h>
+#include <unistd.h>
+
+/* TODO(ophestra): remove after 05ce67fea99ca09cd4b6625cff7aec9cc222dd5a reaches a release */
+#include <sys/syscall.h>
+
+#include "fuse-operations.h"
+
+/* MUST_TRANSLATE_PATHNAME translates a userspace pathname to a relative pathname;
+ * the resulting address points to a constant string or part of pathname, it is never heap allocated. */
+#define MUST_TRANSLATE_PATHNAME(pathname) \
+    do {                                  \
+        if (pathname == NULL)             \
+            return -EINVAL;               \
+        while (*pathname == '/')          \
+            pathname++;                   \
+        if (*pathname == '\0')            \
+            pathname = ".";               \
+    } while (0)
+
+/* GET_CONTEXT_PRIV obtains fuse context and private data for the calling thread. */
+#define GET_CONTEXT_PRIV(ctx, priv) \
+    do {                            \
+        ctx = fuse_get_context();   \
+        priv = ctx->private_data;   \
+    } while (0)
+
+/* impl_getattr modifies a struct stat from the kernel to present to userspace;
+ * impl_getattr returns a negative errno style error code. */
+static int impl_getattr(struct fuse_context *ctx, struct stat *statbuf) {
+    /* allowlist of permitted types */
+    if (!S_ISDIR(statbuf->st_mode) && !S_ISREG(statbuf->st_mode) && !S_ISLNK(statbuf->st_mode)) {
+        return -ENOTRECOVERABLE; /* returning an errno causes all operations on the file to return EIO */
+    }
+
+#define OVERRIDE_PERM(v) (statbuf->st_mode & ~0777) | (v & 0777)
+    if (S_ISDIR(statbuf->st_mode))
+        statbuf->st_mode = OVERRIDE_PERM(SHAREFS_PERM_DIR);
+    else if (S_ISREG(statbuf->st_mode))
+        statbuf->st_mode = OVERRIDE_PERM(SHAREFS_PERM_REG);
+    else
+        statbuf->st_mode = 0; /* should always be symlink in this case */
+
+    statbuf->st_uid = ctx->uid;
+    statbuf->st_gid = SHAREFS_MEDIA_RW_ID;
+    statbuf->st_ctim = statbuf->st_mtim;
+    statbuf->st_nlink = 1;
+    return 0;
+}
+
+/* fuse_operations implementation */
+
+int sharefs_getattr(const char *pathname, struct stat *statbuf, struct fuse_file_info *fi) {
+    struct fuse_context *ctx;
+    struct sharefs_private *priv;
+    GET_CONTEXT_PRIV(ctx, priv);
+    MUST_TRANSLATE_PATHNAME(pathname);
+
+    (void)fi;
+
+    if (fstatat(priv->dirfd, pathname, statbuf, AT_SYMLINK_NOFOLLOW) == -1)
+        return -errno;
+    return impl_getattr(ctx, statbuf);
+}
+
+int sharefs_readdir(const char *pathname, void *buf, fuse_fill_dir_t filler, off_t offset, struct fuse_file_info *fi, enum fuse_readdir_flags flags) {
+    int fd;
+    DIR *dp;
+    struct stat st;
+    int ret = 0;
+    struct dirent *de;
+
+    struct fuse_context *ctx;
+    struct sharefs_private *priv;
+    GET_CONTEXT_PRIV(ctx, priv);
+    MUST_TRANSLATE_PATHNAME(pathname);
+
+    (void)offset;
+    (void)fi;
+
+    if ((fd = openat(priv->dirfd, pathname, O_RDONLY | O_DIRECTORY | O_CLOEXEC)) == -1)
+        return -errno;
+    if ((dp = fdopendir(fd)) == NULL) {
+        close(fd);
+        return -errno;
+    }
+
+    errno = 0; /* for the next readdir call */
+    while ((de = readdir(dp)) != NULL) {
+        if (flags & FUSE_READDIR_PLUS) {
+            if (fstatat(dirfd(dp), de->d_name, &st, AT_SYMLINK_NOFOLLOW) == -1) {
+                ret = -errno;
+                break;
+            }
+
+            if ((ret = impl_getattr(ctx, &st)) < 0)
+                break;
+
+            errno = 0;
+            ret = filler(buf, de->d_name, &st, 0, FUSE_FILL_DIR_PLUS);
+        } else
+            ret = filler(buf, de->d_name, NULL, 0, 0);
+
+        if (ret != 0) {
+            ret = errno != 0 ? -errno : -EIO; /* filler */
+            break;
+        }
+
+        errno = 0; /* for the next readdir call */
+    }
+    if (ret == 0 && errno != 0)
+        ret = -errno; /* readdir */
+
+    closedir(dp);
+    return ret;
+}
+
+int sharefs_mkdir(const char *pathname, mode_t mode) {
+    struct fuse_context *ctx;
+    struct sharefs_private *priv;
+    GET_CONTEXT_PRIV(ctx, priv);
+    MUST_TRANSLATE_PATHNAME(pathname);
+
+    (void)mode;
+
+    if (mkdirat(priv->dirfd, pathname, SHAREFS_PERM_DIR) == -1)
+        return -errno;
+    return 0;
+}
+
+int sharefs_unlink(const char *pathname) {
+    struct fuse_context *ctx;
+    struct sharefs_private *priv;
+    GET_CONTEXT_PRIV(ctx, priv);
+    MUST_TRANSLATE_PATHNAME(pathname);
+
+    if (unlinkat(priv->dirfd, pathname, 0) == -1)
+        return -errno;
+    return 0;
+}
+
+int sharefs_rmdir(const char *pathname) {
+    struct fuse_context *ctx;
+    struct sharefs_private *priv;
+    GET_CONTEXT_PRIV(ctx, priv);
+    MUST_TRANSLATE_PATHNAME(pathname);
+
+    if (unlinkat(priv->dirfd, pathname, AT_REMOVEDIR) == -1)
+        return -errno;
+    return 0;
+}
+
+int sharefs_rename(const char *oldpath, const char *newpath, unsigned int flags) {
+    struct fuse_context *ctx;
+    struct sharefs_private *priv;
+    GET_CONTEXT_PRIV(ctx, priv);
+    MUST_TRANSLATE_PATHNAME(oldpath);
+    MUST_TRANSLATE_PATHNAME(newpath);
+
+    /* TODO(ophestra): replace with wrapper after 05ce67fea99ca09cd4b6625cff7aec9cc222dd5a reaches a release */
+    if (syscall(__NR_renameat2, priv->dirfd, oldpath, priv->dirfd, newpath, flags) == -1)
+        return -errno;
+    return 0;
+}
+
+int sharefs_truncate(const char *pathname, off_t length, struct fuse_file_info *fi) {
+    int fd;
+    int ret;
+
+    struct fuse_context *ctx;
+    struct sharefs_private *priv;
+    GET_CONTEXT_PRIV(ctx, priv);
+    MUST_TRANSLATE_PATHNAME(pathname);
+
+    (void)fi;
+
+    if ((fd = openat(priv->dirfd, pathname, O_WRONLY | O_CLOEXEC)) == -1)
+        return -errno;
+    if ((ret = ftruncate(fd, length)) == -1)
+        ret = -errno;
+    close(fd);
+    return ret;
+}
+
+int sharefs_utimens(const char *pathname, const struct timespec times[2], struct fuse_file_info *fi) {
+    struct fuse_context *ctx;
+    struct sharefs_private *priv;
+    GET_CONTEXT_PRIV(ctx, priv);
+    MUST_TRANSLATE_PATHNAME(pathname);
+
+    (void)fi;
+
+    if (utimensat(priv->dirfd, pathname, times, AT_SYMLINK_NOFOLLOW) == -1)
+        return -errno;
+    return 0;
+}
+
+int sharefs_create(const char *pathname, mode_t mode, struct fuse_file_info *fi) {
+    int fd;
+
+    struct fuse_context *ctx;
+    struct sharefs_private *priv;
+    GET_CONTEXT_PRIV(ctx, priv);
+    MUST_TRANSLATE_PATHNAME(pathname);
+
+    (void)mode;
+
+    if ((fd = openat(priv->dirfd, pathname, fi->flags & ~SHAREFS_FORBIDDEN_FLAGS, SHAREFS_PERM_REG)) == -1)
+        return -errno;
+    fi->fh = fd;
+    return 0;
+}
+
+int sharefs_open(const char *pathname, struct fuse_file_info *fi) {
+    int fd;
+
+    struct fuse_context *ctx;
+    struct sharefs_private *priv;
+    GET_CONTEXT_PRIV(ctx, priv);
+    MUST_TRANSLATE_PATHNAME(pathname);
+
+    if ((fd = openat(priv->dirfd, pathname, fi->flags & ~SHAREFS_FORBIDDEN_FLAGS)) == -1)
+        return -errno;
+    fi->fh = fd;
+    return 0;
+}
+
+int sharefs_read(const char *pathname, char *buf, size_t count, off_t offset, struct fuse_file_info *fi) {
+    int ret;
+
+    (void)pathname;
+
+    if ((ret = pread(fi->fh, buf, count, offset)) == -1)
+        return -errno;
+    return ret;
+}
+
+int sharefs_write(const char *pathname, const char *buf, size_t count, off_t offset, struct fuse_file_info *fi) {
+    int ret;
+
+    (void)pathname;
+
+    if ((ret = pwrite(fi->fh, buf, count, offset)) == -1)
+        return -errno;
+    return ret;
+}
+
+int sharefs_statfs(const char *pathname, struct statvfs *statbuf) {
+    int fd;
+    int ret;
+
+    struct fuse_context *ctx;
+    struct sharefs_private *priv;
+    GET_CONTEXT_PRIV(ctx, priv);
+    MUST_TRANSLATE_PATHNAME(pathname);
+
+    if ((fd = openat(priv->dirfd, pathname, O_RDONLY | O_CLOEXEC)) == -1)
+        return -errno;
+    if ((ret = fstatvfs(fd, statbuf)) == -1)
+        ret = -errno;
+    close(fd);
+    return ret;
+}
+
+int sharefs_release(const char *pathname, struct fuse_file_info *fi) {
+    (void)pathname;
+
+    return close(fi->fh);
+}
+
+int sharefs_fsync(const char *pathname, int datasync, struct fuse_file_info *fi) {
+    (void)pathname;
+
+    if (datasync ? fdatasync(fi->fh) : fsync(fi->fh) == -1)
+        return -errno;
+    return 0;
+}

cmd/sharefs/fuse-operations.h

@@ -0,0 +1,34 @@
+#define FUSE_USE_VERSION FUSE_MAKE_VERSION(3, 12)
+#include <fuse.h>
+#include <fuse_lowlevel.h> /* for fuse_cmdline_help */
+
+#if (FUSE_VERSION < FUSE_MAKE_VERSION(3, 12))
+#error This package requires libfuse >= v3.12
+#endif
+
+#define SHAREFS_MEDIA_RW_ID (1 << 10) - 1 /* owning gid presented to userspace */
+#define SHAREFS_PERM_DIR 0700             /* permission bits for directories presented to userspace */
+#define SHAREFS_PERM_REG 0600             /* permission bits for regular files presented to userspace */
+#define SHAREFS_FORBIDDEN_FLAGS O_DIRECT  /* these open flags are cleared unconditionally */
+
+/* sharefs_private is populated by sharefs_init and contains process-wide context */
+struct sharefs_private {
+    int dirfd;       /* source dirfd opened during sharefs_init */
+    uintptr_t setup; /* cgo handle of opaque setup state */
+};
+
+int sharefs_getattr(const char *pathname, struct stat *statbuf, struct fuse_file_info *fi);
+int sharefs_readdir(const char *pathname, void *buf, fuse_fill_dir_t filler, off_t offset, struct fuse_file_info *fi, enum fuse_readdir_flags flags);
+int sharefs_mkdir(const char *pathname, mode_t mode);
+int sharefs_unlink(const char *pathname);
+int sharefs_rmdir(const char *pathname);
+int sharefs_rename(const char *oldpath, const char *newpath, unsigned int flags);
+int sharefs_truncate(const char *pathname, off_t length, struct fuse_file_info *fi);
+int sharefs_utimens(const char *pathname, const struct timespec times[2], struct fuse_file_info *fi);
+int sharefs_create(const char *pathname, mode_t mode, struct fuse_file_info *fi);
+int sharefs_open(const char *pathname, struct fuse_file_info *fi);
+int sharefs_read(const char *pathname, char *buf, size_t count, off_t offset, struct fuse_file_info *fi);
+int sharefs_write(const char *pathname, const char *buf, size_t count, off_t offset, struct fuse_file_info *fi);
+int sharefs_statfs(const char *pathname, struct statvfs *statbuf);
+int sharefs_release(const char *pathname, struct fuse_file_info *fi);
+int sharefs_fsync(const char *pathname, int datasync, struct fuse_file_info *fi);

cmd/sharefs/fuse.go

@@ -0,0 +1,556 @@
+package main
+
+/*
+#cgo pkg-config: --static fuse3
+
+#include "fuse-operations.h"
+#include <stdlib.h>
+#include <string.h>
+
+extern void *sharefs_init(struct fuse_conn_info *conn, struct fuse_config *cfg);
+extern void sharefs_destroy(void *private_data);
+
+typedef void (*closure)();
+static inline struct fuse_opt _FUSE_OPT_END() { return (struct fuse_opt)FUSE_OPT_END; };
+*/
+import "C"
+import (
+	"context"
+	"encoding/gob"
+	"errors"
+	"fmt"
+	"io"
+	"log"
+	"os"
+	"os/exec"
+	"os/signal"
+	"path"
+	"runtime"
+	"runtime/cgo"
+	"strconv"
+	"syscall"
+	"unsafe"
+
+	"hakurei.app/container"
+	"hakurei.app/container/check"
+	"hakurei.app/container/std"
+	"hakurei.app/hst"
+	"hakurei.app/internal/helper/proc"
+	"hakurei.app/internal/info"
+	"hakurei.app/message"
+)
+
+type (
+	// closure represents a C function pointer.
+	closure = C.closure
+
+	// fuseArgs represents the fuse_args structure.
+	fuseArgs = C.struct_fuse_args
+
+	// setupState holds state used for setup. Its cgo handle is included in
+	// sharefs_private and considered opaque to non-setup callbacks.
+	setupState struct {
+		// Whether sharefs_init failed.
+		initFailed bool
+
+		// Whether to create source directory as root.
+		mkdir bool
+
+		// Open file descriptor to fuse.
+		Fuse int
+
+		// Pathname to open for dirfd.
+		Source *check.Absolute
+		// New uid and gid to set by sharefs_init when starting as root.
+		Setuid, Setgid int
+	}
+)
+
+func init() { gob.Register(new(setupState)) }
+
+// destroySetup invalidates the setup [cgo.Handle] in a sharefs_private structure.
+func destroySetup(private_data unsafe.Pointer) (ok bool) {
+	if private_data == nil {
+		return false
+	}
+	priv := (*C.struct_sharefs_private)(private_data)
+
+	if h := cgo.Handle(priv.setup); h != 0 {
+		priv.setup = 0
+		h.Delete()
+		ok = true
+	}
+	return
+}
+
+//export sharefs_init
+func sharefs_init(_ *C.struct_fuse_conn_info, cfg *C.struct_fuse_config) unsafe.Pointer {
+	ctx := C.fuse_get_context()
+	priv := (*C.struct_sharefs_private)(ctx.private_data)
+	setup := cgo.Handle(priv.setup).Value().(*setupState)
+
+	if os.Geteuid() == 0 {
+		log.Println("filesystem daemon must not run as root")
+		goto fail
+	}
+
+	cfg.use_ino = C.true
+	cfg.direct_io = C.false
+	// getattr is context-dependent
+	cfg.attr_timeout = 0
+	cfg.entry_timeout = 0
+	cfg.negative_timeout = 0
+
+	// all future filesystem operations happen through this dirfd
+	if fd, err := syscall.Open(setup.Source.String(), syscall.O_DIRECTORY|syscall.O_RDONLY|syscall.O_CLOEXEC, 0); err != nil {
+		log.Printf("cannot open %q: %v", setup.Source, err)
+		goto fail
+	} else if err = syscall.Fchdir(fd); err != nil {
+		_ = syscall.Close(fd)
+		log.Printf("cannot enter %q: %s", setup.Source, err)
+		goto fail
+	} else {
+		priv.dirfd = C.int(fd)
+	}
+
+	return ctx.private_data
+
+fail:
+	setup.initFailed = true
+	C.fuse_exit(ctx.fuse)
+	return nil
+}
+
+//export sharefs_destroy
+func sharefs_destroy(private_data unsafe.Pointer) {
+	if private_data != nil {
+		destroySetup(private_data)
+		priv := (*C.struct_sharefs_private)(private_data)
+
+		if err := syscall.Close(int(priv.dirfd)); err != nil {
+			log.Printf("cannot close source directory: %v", err)
+		}
+	}
+}
+
+// showHelp prints the help message.
+func showHelp(args *fuseArgs) {
+	executableName := sharefsName
+	if args.argc > 0 {
+		executableName = path.Base(C.GoString(*args.argv))
+	} else if name, err := os.Executable(); err == nil {
+		executableName = path.Base(name)
+	}
+
+	fmt.Printf("usage: %s [options] <mountpoint>\n\n", executableName)
+
+	fmt.Println("Filesystem options:")
+	fmt.Println("    -o source=/data/media  source directory to be mounted")
+	fmt.Println("    -o setuid=1023         uid to run as when starting as root")
+	fmt.Println("    -o setgid=1023         gid to run as when starting as root")
+
+	fmt.Println("\nFUSE options:")
+	C.fuse_cmdline_help()
+	C.fuse_lib_help(args)
+}
+
+// parseOpts parses fuse options via fuse_opt_parse.
+func parseOpts(args *fuseArgs, setup *setupState, log *log.Logger) (ok bool) {
+	var unsafeOpts struct {
+		// Pathname to writable source directory.
+		source *C.char
+
+		// Whether to create source directory as root.
+		mkdir C.int
+
+		// Decimal string representation of uid to set when running as root.
+		setuid *C.char
+		// Decimal string representation of gid to set when running as root.
+		setgid *C.char
+
+		// Decimal string representation of open file descriptor to read setupState from.
+		// This is an internal detail for containerisation and must not be specified directly.
+		setup *C.char
+	}
+
+	if C.fuse_opt_parse(args, unsafe.Pointer(&unsafeOpts), &[]C.struct_fuse_opt{
+		{templ: C.CString("source=%s"), offset: C.ulong(unsafe.Offsetof(unsafeOpts.source)), value: 0},
+		{templ: C.CString("mkdir"), offset: C.ulong(unsafe.Offsetof(unsafeOpts.mkdir)), value: 1},
+		{templ: C.CString("setuid=%s"), offset: C.ulong(unsafe.Offsetof(unsafeOpts.setuid)), value: 0},
+		{templ: C.CString("setgid=%s"), offset: C.ulong(unsafe.Offsetof(unsafeOpts.setgid)), value: 0},
+
+		{templ: C.CString("setup=%s"), offset: C.ulong(unsafe.Offsetof(unsafeOpts.setup)), value: 0},
+
+		C._FUSE_OPT_END(),
+	}[0], nil) == -1 {
+		return false
+	}
+
+	if unsafeOpts.source != nil {
+		defer C.free(unsafe.Pointer(unsafeOpts.source))
+	}
+	if unsafeOpts.setuid != nil {
+		defer C.free(unsafe.Pointer(unsafeOpts.setuid))
+	}
+	if unsafeOpts.setgid != nil {
+		defer C.free(unsafe.Pointer(unsafeOpts.setgid))
+	}
+
+	if unsafeOpts.setup != nil {
+		defer C.free(unsafe.Pointer(unsafeOpts.setup))
+
+		if v, err := strconv.Atoi(C.GoString(unsafeOpts.setup)); err != nil || v < 3 {
+			log.Println("invalid value for option setup")
+			return false
+		} else {
+			r := os.NewFile(uintptr(v), "setup")
+			defer func() {
+				if err = r.Close(); err != nil {
+					log.Println(err)
+				}
+			}()
+			if err = gob.NewDecoder(r).Decode(setup); err != nil {
+				log.Println(err)
+				return false
+			}
+		}
+		if setup.Fuse < 3 {
+			log.Println("invalid file descriptor", setup.Fuse)
+			return false
+		}
+		return true
+	}
+
+	if unsafeOpts.source == nil {
+		showHelp(args)
+		return false
+	} else if a, err := check.NewAbs(C.GoString(unsafeOpts.source)); err != nil {
+		log.Println(err)
+		return false
+	} else {
+		setup.Source = a
+	}
+	setup.mkdir = unsafeOpts.mkdir != 0
+
+	if unsafeOpts.setuid == nil {
+		setup.Setuid = -1
+	} else if v, err := strconv.Atoi(C.GoString(unsafeOpts.setuid)); err != nil || v <= 0 {
+		log.Println("invalid value for option setuid")
+		return false
+	} else {
+		setup.Setuid = v
+	}
+	if unsafeOpts.setgid == nil {
+		setup.Setgid = -1
+	} else if v, err := strconv.Atoi(C.GoString(unsafeOpts.setgid)); err != nil || v <= 0 {
+		log.Println("invalid value for option setgid")
+		return false
+	} else {
+		setup.Setgid = v
+	}
+
+	return true
+}
+
+// copyArgs returns a heap allocated copy of an argument slice in fuse_args representation.
+func copyArgs(s ...string) fuseArgs {
+	if len(s) == 0 {
+		return fuseArgs{argc: 0, argv: nil, allocated: 0}
+	}
+	args := unsafe.Slice((**C.char)(C.malloc(C.size_t(uintptr(len(s))*unsafe.Sizeof(s[0])))), len(s))
+	for i, arg := range s {
+		args[i] = C.CString(arg)
+	}
+	return fuseArgs{argc: C.int(len(s)), argv: &args[0], allocated: 1}
+}
+
+// freeArgs frees the contents of argument list.
+func freeArgs(args *fuseArgs) { C.fuse_opt_free_args(args) }
+
+// unsafeAddArgument adds an argument to fuseArgs via fuse_opt_add_arg.
+// The last byte of arg must be 0.
+func unsafeAddArgument(args *fuseArgs, arg string) {
+	C.fuse_opt_add_arg(args, (*C.char)(unsafe.Pointer(unsafe.StringData(arg))))
+}
+
+func _main(s ...string) (exitCode int) {
+	msg := message.New(log.Default())
+	container.TryArgv0(msg)
+	runtime.LockOSThread()
+
+	// don't mask creation mode, kernel already did that
+	syscall.Umask(0)
+
+	var pinner runtime.Pinner
+	defer pinner.Unpin()
+
+	args := copyArgs(s...)
+	defer freeArgs(&args)
+
+	// this causes the kernel to enforce access control based on
+	// struct stat populated by sharefs_getattr
+	unsafeAddArgument(&args, "-odefault_permissions\x00")
+
+	var priv C.struct_sharefs_private
+	pinner.Pin(&priv)
+	var setup setupState
+	priv.setup = C.uintptr_t(cgo.NewHandle(&setup))
+	defer destroySetup(unsafe.Pointer(&priv))
+
+	var opts C.struct_fuse_cmdline_opts
+	if C.fuse_parse_cmdline(&args, &opts) != 0 {
+		return 1
+	}
+	if opts.mountpoint != nil {
+		defer C.free(unsafe.Pointer(opts.mountpoint))
+	}
+
+	if opts.show_version != 0 {
+		fmt.Println("hakurei version", info.Version())
+		fmt.Println("FUSE library version", C.GoString(C.fuse_pkgversion()))
+		C.fuse_lowlevel_version()
+		return 0
+	}
+
+	if opts.show_help != 0 {
+		showHelp(&args)
+		return 0
+	} else if opts.mountpoint == nil {
+		log.Println("no mountpoint specified")
+		return 2
+	} else {
+		// hack to keep fuse_parse_cmdline happy in the container
+		mountpoint := C.GoString(opts.mountpoint)
+		pathnameArg := -1
+		for i, arg := range s {
+			if arg == mountpoint {
+				pathnameArg = i
+				break
+			}
+		}
+		if pathnameArg < 0 {
+			log.Println("mountpoint must be absolute")
+			return 2
+		}
+		s[pathnameArg] = container.Nonexistent
+	}
+
+	if !parseOpts(&args, &setup, msg.GetLogger()) {
+		return 1
+	}
+	asRoot := os.Geteuid() == 0
+
+	if asRoot {
+		if setup.Setuid <= 0 || setup.Setgid <= 0 {
+			log.Println("setuid and setgid must not be 0")
+			return 1
+		}
+
+		if setup.Fuse >= 3 {
+			log.Println("filesystem daemon must not run as root")
+			return 1
+		}
+
+		if setup.mkdir {
+			if err := os.MkdirAll(setup.Source.String(), 0700); err != nil {
+				if !errors.Is(err, os.ErrExist) {
+					log.Println(err)
+					return 1
+				}
+				// skip setup for existing source directory
+			} else if err = os.Chown(setup.Source.String(), setup.Setuid, setup.Setgid); err != nil {
+				log.Println(err)
+				return 1
+			}
+		}
+	} else if setup.Fuse < 3 && (setup.Setuid > 0 || setup.Setgid > 0) {
+		log.Println("setuid and setgid has no effect when not starting as root")
+		return 1
+	} else if setup.mkdir {
+		log.Println("mkdir has no effect when not starting as root")
+		return 1
+	}
+
+	op := C.struct_fuse_operations{
+		init:    closure(C.sharefs_init),
+		destroy: closure(C.sharefs_destroy),
+
+		// implemented in fuse-helper.c
+		getattr:  closure(C.sharefs_getattr),
+		readdir:  closure(C.sharefs_readdir),
+		mkdir:    closure(C.sharefs_mkdir),
+		unlink:   closure(C.sharefs_unlink),
+		rmdir:    closure(C.sharefs_rmdir),
+		rename:   closure(C.sharefs_rename),
+		truncate: closure(C.sharefs_truncate),
+		utimens:  closure(C.sharefs_utimens),
+		create:   closure(C.sharefs_create),
+		open:     closure(C.sharefs_open),
+		read:     closure(C.sharefs_read),
+		write:    closure(C.sharefs_write),
+		statfs:   closure(C.sharefs_statfs),
+		release:  closure(C.sharefs_release),
+		fsync:    closure(C.sharefs_fsync),
+	}
+
+	fuse := C.fuse_new_fn(&args, &op, C.size_t(unsafe.Sizeof(op)), unsafe.Pointer(&priv))
+	if fuse == nil {
+		return 3
+	}
+	defer C.fuse_destroy(fuse)
+	se := C.fuse_get_session(fuse)
+
+	if setup.Fuse < 3 {
+		// unconfined, set up mount point and container
+		if C.fuse_mount(fuse, opts.mountpoint) != 0 {
+			return 4
+		}
+		// unmounted by initial process
+		defer func() {
+			if exitCode == 5 {
+				C.fuse_unmount(fuse)
+			}
+		}()
+
+		if asRoot {
+			if err := syscall.Setresgid(setup.Setgid, setup.Setgid, setup.Setgid); err != nil {
+				log.Printf("cannot set gid: %v", err)
+				return 5
+			}
+			if err := syscall.Setgroups(nil); err != nil {
+				log.Printf("cannot set supplementary groups: %v", err)
+				return 5
+			}
+			if err := syscall.Setresuid(setup.Setuid, setup.Setuid, setup.Setuid); err != nil {
+				log.Printf("cannot set uid: %v", err)
+				return 5
+			}
+		}
+
+		msg.SwapVerbose(opts.debug != 0)
+		ctx := context.Background()
+		if opts.foreground != 0 {
+			c, cancel := signal.NotifyContext(ctx, syscall.SIGINT, syscall.SIGTERM)
+			defer cancel()
+			ctx = c
+		}
+		z := container.New(ctx, msg)
+		z.AllowOrphan = opts.foreground == 0
+		z.Env = os.Environ()
+
+		// keep fuse_parse_cmdline happy in the container
+		z.Tmpfs(check.MustAbs(container.Nonexistent), 1<<10, 0755)
+
+		if a, err := check.NewAbs(container.MustExecutable(msg)); err != nil {
+			log.Println(err)
+			return 5
+		} else {
+			z.Path = a
+		}
+		z.Args = s
+		z.ForwardCancel = true
+		z.SeccompPresets |= std.PresetStrict
+		z.ParentPerm = 0700
+		z.Bind(setup.Source, setup.Source, std.BindWritable)
+		if !z.AllowOrphan {
+			z.WaitDelay = hst.WaitDelayMax
+			z.Stdin, z.Stdout, z.Stderr = os.Stdin, os.Stdout, os.Stderr
+		}
+		z.Bind(z.Path, z.Path, 0)
+		setup.Fuse = int(proc.ExtraFileSlice(&z.ExtraFiles, os.NewFile(uintptr(C.fuse_session_fd(se)), "fuse")))
+
+		var setupWriter io.WriteCloser
+		if fd, w, err := container.Setup(&z.ExtraFiles); err != nil {
+			log.Println(err)
+			return 5
+		} else {
+			z.Args = append(z.Args, "-osetup="+strconv.Itoa(fd))
+			setupWriter = w
+		}
+
+		if err := z.Start(); err != nil {
+			if m, ok := message.GetMessage(err); ok {
+				log.Println(m)
+			} else {
+				log.Println(err)
+			}
+			return 5
+		}
+		if err := z.Serve(); err != nil {
+			if m, ok := message.GetMessage(err); ok {
+				log.Println(m)
+			} else {
+				log.Println(err)
+			}
+			return 5
+		}
+
+		if err := gob.NewEncoder(setupWriter).Encode(&setup); err != nil {
+			log.Println(err)
+			return 5
+		} else if err = setupWriter.Close(); err != nil {
+			log.Println(err)
+		}
+
+		if !z.AllowOrphan {
+			if err := z.Wait(); err != nil {
+				var exitError *exec.ExitError
+				if !errors.As(err, &exitError) || exitError == nil {
+					log.Println(err)
+					return 5
+				}
+				switch code := exitError.ExitCode(); syscall.Signal(code & 0x7f) {
+				case syscall.SIGINT:
+				case syscall.SIGTERM:
+
+				default:
+					return code
+				}
+			}
+		}
+		return 0
+	} else { // confined
+		C.free(unsafe.Pointer(opts.mountpoint))
+		// must be heap allocated
+		opts.mountpoint = C.CString("/dev/fd/" + strconv.Itoa(setup.Fuse))
+
+		if err := os.Chdir("/"); err != nil {
+			log.Println(err)
+		}
+	}
+
+	if C.fuse_mount(fuse, opts.mountpoint) != 0 {
+		return 4
+	}
+	defer C.fuse_unmount(fuse)
+
+	if C.fuse_set_signal_handlers(se) != 0 {
+		return 6
+	}
+	defer C.fuse_remove_signal_handlers(se)
+
+	if opts.singlethread != 0 {
+		if C.fuse_loop(fuse) != 0 {
+			return 8
+		}
+	} else {
+		loopConfig := C.fuse_loop_cfg_create()
+		if loopConfig == nil {
+			return 7
+		}
+		defer C.fuse_loop_cfg_destroy(loopConfig)
+
+		C.fuse_loop_cfg_set_clone_fd(loopConfig, C.uint(opts.clone_fd))
+
+		C.fuse_loop_cfg_set_idle_threads(loopConfig, opts.max_idle_threads)
+		C.fuse_loop_cfg_set_max_threads(loopConfig, opts.max_threads)
+		if C.fuse_loop_mt(fuse, loopConfig) != 0 {
+			return 8
+		}
+	}
+
+	if setup.initFailed {
+		return 1
+	}
+	return 0
+}

cmd/sharefs/fuse_test.go

@@ -0,0 +1,113 @@
+package main
+
+import (
+	"bytes"
+	"log"
+	"reflect"
+	"testing"
+
+	"hakurei.app/container/check"
+)
+
+func TestParseOpts(t *testing.T) {
+	t.Parallel()
+
+	testCases := []struct {
+		name    string
+		args    []string
+		want    setupState
+		wantLog string
+		wantOk  bool
+	}{
+		{"zero length", []string{}, setupState{}, "", false},
+
+		{"not absolute", []string{"sharefs",
+			"-o", "source=nonexistent",
+			"-o", "setuid=1023",
+			"-o", "setgid=1023",
+		}, setupState{}, "sharefs: path \"nonexistent\" is not absolute\n", false},
+
+		{"not specified", []string{"sharefs",
+			"-o", "setuid=1023",
+			"-o", "setgid=1023",
+		}, setupState{}, "", false},
+
+		{"invalid setuid", []string{"sharefs",
+			"-o", "source=/proc/nonexistent",
+			"-o", "setuid=ff",
+			"-o", "setgid=1023",
+		}, setupState{
+			Source: check.MustAbs("/proc/nonexistent"),
+		}, "sharefs: invalid value for option setuid\n", false},
+
+		{"invalid setgid", []string{"sharefs",
+			"-o", "source=/proc/nonexistent",
+			"-o", "setuid=1023",
+			"-o", "setgid=ff",
+		}, setupState{
+			Source: check.MustAbs("/proc/nonexistent"),
+			Setuid: 1023,
+		}, "sharefs: invalid value for option setgid\n", false},
+
+		{"simple", []string{"sharefs",
+			"-o", "source=/proc/nonexistent",
+		}, setupState{
+			Source: check.MustAbs("/proc/nonexistent"),
+			Setuid: -1,
+			Setgid: -1,
+		}, "", true},
+
+		{"root", []string{"sharefs",
+			"-o", "source=/proc/nonexistent",
+			"-o", "setuid=1023",
+			"-o", "setgid=1023",
+		}, setupState{
+			Source: check.MustAbs("/proc/nonexistent"),
+			Setuid: 1023,
+			Setgid: 1023,
+		}, "", true},
+
+		{"setuid", []string{"sharefs",
+			"-o", "source=/proc/nonexistent",
+			"-o", "setuid=1023",
+		}, setupState{
+			Source: check.MustAbs("/proc/nonexistent"),
+			Setuid: 1023,
+			Setgid: -1,
+		}, "", true},
+
+		{"setgid", []string{"sharefs",
+			"-o", "source=/proc/nonexistent",
+			"-o", "setgid=1023",
+		}, setupState{
+			Source: check.MustAbs("/proc/nonexistent"),
+			Setuid: -1,
+			Setgid: 1023,
+		}, "", true},
+	}
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Parallel()
+
+			var (
+				got setupState
+				buf bytes.Buffer
+			)
+			args := copyArgs(tc.args...)
+			defer freeArgs(&args)
+			unsafeAddArgument(&args, "-odefault_permissions\x00")
+
+			if ok := parseOpts(&args, &got, log.New(&buf, "sharefs: ", 0)); ok != tc.wantOk {
+				t.Errorf("parseOpts: ok = %v, want %v", ok, tc.wantOk)
+			}
+
+			if !reflect.DeepEqual(&got, &tc.want) {
+				t.Errorf("parseOpts: setup = %#v, want %#v", got, tc.want)
+			}
+
+			if buf.String() != tc.wantLog {
+				t.Errorf("parseOpts: log =\n%s\nwant\n%s", buf.String(), tc.wantLog)
+			}
+		})
+	}
+}

cmd/sharefs/main.go

@@ -0,0 +1,31 @@
+package main
+
+import (
+	"log"
+	"os"
+	"slices"
+)
+
+// sharefsName is the prefix used by log.std in the sharefs process.
+const sharefsName = "sharefs"
+
+// handleMountArgs returns an alternative, libfuse-compatible args slice for
+// args passed by mount -t fuse.sharefs [options] sharefs <mountpoint>.
+//
+// In this case, args always has a length of 5 with index 0 being what comes
+// after "fuse." in the filesystem type, 1 is the uninterpreted string passed
+// to mount (sharefsName is used as the magic string to enable this hack),
+// 2 is passed through to libfuse as mountpoint, and 3 is always "-o".
+func handleMountArgs(args []string) []string {
+	if len(args) == 5 && args[1] == sharefsName && args[3] == "-o" {
+		return []string{sharefsName, args[2], "-o", args[4]}
+	}
+	return slices.Clone(args)
+}
+
+func main() {
+	log.SetFlags(0)
+	log.SetPrefix(sharefsName + ": ")
+
+	os.Exit(_main(handleMountArgs(os.Args)...))
+}

cmd/sharefs/main_test.go

@@ -0,0 +1,29 @@
+package main
+
+import (
+	"slices"
+	"testing"
+)
+
+func TestHandleMountArgs(t *testing.T) {
+	t.Parallel()
+
+	testCases := []struct {
+		name string
+		args []string
+		want []string
+	}{
+		{"nil", nil, nil},
+		{"passthrough", []string{"sharefs", "-V"}, []string{"sharefs", "-V"}},
+		{"replace", []string{"/sbin/sharefs", "sharefs", "/sdcard", "-o", "rw"}, []string{"sharefs", "/sdcard", "-o", "rw"}},
+	}
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Parallel()
+
+			if got := handleMountArgs(tc.args); !slices.Equal(got, tc.want) {
+				t.Errorf("handleMountArgs: %q, want %q", got, tc.want)
+			}
+		})
+	}
+}

cmd/sharefs/test/configuration.nix

@@ -0,0 +1,41 @@
+{ pkgs, ... }:
+{
+  users.users = {
+    alice = {
+      isNormalUser = true;
+      description = "Alice Foobar";
+      password = "foobar";
+      uid = 1000;
+    };
+  };
+
+  home-manager.users.alice.home.stateVersion = "24.11";
+
+  # Automatically login on tty1 as a normal user:
+  services.getty.autologinUser = "alice";
+
+  environment = {
+    # For benchmarking sharefs:
+    systemPackages = [ pkgs.fsmark ];
+  };
+
+  virtualisation = {
+    diskSize = 6 * 1024;
+
+    qemu.options = [
+      # Increase test performance:
+      "-smp 8"
+    ];
+  };
+
+  environment.hakurei = rec {
+    enable = true;
+    stateDir = "/var/lib/hakurei";
+    sharefs.source = "${stateDir}/sdcard";
+    users.alice = 0;
+
+    extraHomeConfig = {
+      home.stateVersion = "23.05";
+    };
+  };
+}

cmd/sharefs/test/default.nix

@@ -0,0 +1,44 @@
+{
+  testers,
+
+  system,
+  self,
+}:
+testers.nixosTest {
+  name = "sharefs";
+  nodes.machine =
+    { options, pkgs, ... }:
+    let
+      fhs =
+        let
+          hakurei = options.environment.hakurei.package.default;
+        in
+        pkgs.buildFHSEnv {
+          pname = "hakurei-fhs";
+          inherit (hakurei) version;
+          targetPkgs = _: hakurei.targetPkgs;
+          extraOutputsToInstall = [ "dev" ];
+          profile = ''
+            export PKG_CONFIG_PATH="/usr/share/pkgconfig:$PKG_CONFIG_PATH"
+          '';
+        };
+    in
+    {
+      environment.systemPackages = [
+        # For go tests:
+        (pkgs.writeShellScriptBin "sharefs-workload-hakurei-tests" ''
+          cp -r "${self.packages.${system}.hakurei.src}" "/sdcard/hakurei" && cd "/sdcard/hakurei"
+          ${fhs}/bin/hakurei-fhs -c 'CC="clang -O3 -Werror" go test ./...'
+        '')
+      ];
+
+      imports = [
+        ./configuration.nix
+
+        self.nixosModules.hakurei
+        self.inputs.home-manager.nixosModules.home-manager
+      ];
+    };
+
+  testScript = builtins.readFile ./test.py;
+}

cmd/sharefs/test/test.py

@@ -0,0 +1,60 @@
+start_all()
+machine.wait_for_unit("multi-user.target")
+
+# To check sharefs version:
+print(machine.succeed("sharefs -V"))
+
+# Make sure sharefs started:
+machine.wait_for_unit("sdcard.mount")
+
+machine.succeed("mkdir /mnt")
+def check_bad_opts_output(opts, want, source="/etc", privileged=False):
+    output = machine.fail(("" if privileged else "sudo -u alice -i ") + f"sharefs -f -o source={source},{opts} /mnt 2>&1")
+    if output != want:
+        raise Exception(f"unexpected output: {output}")
+
+# Malformed setuid/setgid representation:
+check_bad_opts_output("setuid=ff", "sharefs: invalid value for option setuid\n")
+check_bad_opts_output("setgid=ff", "sharefs: invalid value for option setgid\n")
+
+# Bounds check for setuid/setgid:
+check_bad_opts_output("setuid=0", "sharefs: invalid value for option setuid\n")
+check_bad_opts_output("setgid=0", "sharefs: invalid value for option setgid\n")
+check_bad_opts_output("setuid=-1", "sharefs: invalid value for option setuid\n")
+check_bad_opts_output("setgid=-1", "sharefs: invalid value for option setgid\n")
+
+# Non-root setuid/setgid:
+check_bad_opts_output("setuid=1023", "sharefs: setuid and setgid has no effect when not starting as root\n")
+check_bad_opts_output("setgid=1023", "sharefs: setuid and setgid has no effect when not starting as root\n")
+check_bad_opts_output("setuid=1023,setgid=1023", "sharefs: setuid and setgid has no effect when not starting as root\n")
+check_bad_opts_output("mkdir", "sharefs: mkdir has no effect when not starting as root\n")
+
+# Starting as root without setuid/setgid:
+check_bad_opts_output("allow_other", "sharefs: setuid and setgid must not be 0\n", privileged=True)
+check_bad_opts_output("setuid=1023", "sharefs: setuid and setgid must not be 0\n", privileged=True)
+check_bad_opts_output("setgid=1023", "sharefs: setuid and setgid must not be 0\n", privileged=True)
+
+# Make sure nothing actually got mounted:
+machine.fail("umount /mnt")
+machine.succeed("rmdir /mnt")
+
+# Unprivileged mount/unmount:
+machine.succeed("sudo -u alice -i mkdir /home/alice/{sdcard,persistent}")
+machine.succeed("sudo -u alice -i sharefs -o source=/home/alice/persistent /home/alice/sdcard")
+machine.succeed("sudo -u alice -i touch /home/alice/sdcard/check")
+machine.succeed("sudo -u alice -i umount /home/alice/sdcard")
+machine.succeed("sudo -u alice -i rm /home/alice/persistent/check")
+machine.succeed("sudo -u alice -i rmdir /home/alice/{sdcard,persistent}")
+
+# Benchmark sharefs:
+machine.succeed("fs_mark -v -d /sdcard/fs_mark -l /tmp/fs_log.txt")
+machine.copy_from_vm("/tmp/fs_log.txt", "")
+
+# Check permissions:
+machine.succeed("sudo -u sharefs touch /var/lib/hakurei/sdcard/fs_mark/.check")
+machine.succeed("sudo -u sharefs rm /var/lib/hakurei/sdcard/fs_mark/.check")
+machine.succeed("sudo -u alice rm -rf /sdcard/fs_mark")
+machine.fail("ls /var/lib/hakurei/sdcard/fs_mark")
+
+# Run hakurei tests on sharefs:
+machine.succeed("sudo -u alice -i sharefs-workload-hakurei-tests")

container/autoetc.go

@@ -59,6 +59,7 @@ func (e *AutoEtcOp) apply(state *setupState, k syscallDispatcher) error {
 
 	return nil
 }
+func (e *AutoEtcOp) late(*setupState, syscallDispatcher) error { return nil }
 
 func (e *AutoEtcOp) hostPath() *check.Absolute { return fhs.AbsEtc.Append(e.hostRel()) }
 func (e *AutoEtcOp) hostRel() string           { return ".host/" + e.Prefix }

container/autoroot.go

@@ -69,6 +69,7 @@ func (r *AutoRootOp) apply(state *setupState, k syscallDispatcher) error {
 	}
 	return nil
 }
+func (r *AutoRootOp) late(*setupState, syscallDispatcher) error { return nil }
 
 func (r *AutoRootOp) Is(op Op) bool {
 	vr, ok := op.(*AutoRootOp)

container/autoroot_test.go

@@ -202,7 +202,7 @@ func TestIsAutoRootBindable(t *testing.T) {
 			t.Parallel()
 			var msg message.Msg
 			if tc.log {
-				msg = &kstub{nil, stub.New(t, func(s *stub.Stub[syscallDispatcher]) syscallDispatcher { panic("unreachable") }, stub.Expect{Calls: []stub.Call{
+				msg = &kstub{nil, nil, stub.New(t, func(s *stub.Stub[syscallDispatcher]) syscallDispatcher { panic("unreachable") }, stub.Expect{Calls: []stub.Call{
 					call("verbose", stub.ExpectArgs{[]any{"got unexpected root entry"}}, nil, nil),
 				}})}
 			}

container/check/absolute.go

@@ -56,7 +56,7 @@ func NewAbs(pathname string) (*Absolute, error) {
 // MustAbs calls [NewAbs] and panics on error.
 func MustAbs(pathname string) *Absolute {
 	if a, err := NewAbs(pathname); err != nil {
-		panic(err.Error())
+		panic(err)
 	} else {
 		return a
 	}

container/check/absolute_test.go

@@ -14,8 +14,10 @@ import (
 	. "hakurei.app/container/check"
 )
 
+// unsafeAbs returns check.Absolute on any string value.
+//
 //go:linkname unsafeAbs hakurei.app/container/check.unsafeAbs
-func unsafeAbs(_ string) *Absolute
+func unsafeAbs(pathname string) *Absolute
 
 func TestAbsoluteError(t *testing.T) {
 	t.Parallel()
@@ -82,9 +84,9 @@ func TestNewAbs(t *testing.T) {
 		t.Parallel()
 
 		defer func() {
-			wantPanic := `path "etc" is not absolute`
+			wantPanic := &AbsoluteError{Pathname: "etc"}
 
-			if r := recover(); r != wantPanic {
+			if r := recover(); !reflect.DeepEqual(r, wantPanic) {
 				t.Errorf("MustAbs: panic = %v; want %v", r, wantPanic)
 			}
 		}()

container/container.go

@@ -35,6 +35,8 @@ type (
 	// Container represents a container environment being prepared or run.
 	// None of [Container] methods are safe for concurrent use.
 	Container struct {
+		// Whether the container init should stay alive after its parent terminates.
+		AllowOrphan bool
 		// Cgroup fd, nil to disable.
 		Cgroup *int
 		// ExtraFiles passed through to initial process in the container,
@@ -252,8 +254,7 @@ func (p *Container) Start() error {
 	}
 	p.cmd.Dir = fhs.Root
 	p.cmd.SysProcAttr = &SysProcAttr{
-		Setsid:    !p.RetainSession,
-		Pdeathsig: SIGKILL,
+		Setsid: !p.RetainSession,
 		Cloneflags: CLONE_NEWUSER | CLONE_NEWPID | CLONE_NEWNS |
 			CLONE_NEWIPC | CLONE_NEWUTS | CLONE_NEWCGROUP,
 
@@ -268,6 +269,9 @@ func (p *Container) Start() error {
 
 		UseCgroupFD: p.Cgroup != nil,
 	}
+	if !p.AllowOrphan {
+		p.cmd.SysProcAttr.Pdeathsig = SIGKILL
+	}
 	if p.cmd.SysProcAttr.UseCgroupFD {
 		p.cmd.SysProcAttr.CgroupFD = *p.Cgroup
 	}

container/container_test.go

@@ -21,6 +21,7 @@ import (
 	"hakurei.app/command"
 	"hakurei.app/container"
 	"hakurei.app/container/check"
+	"hakurei.app/container/fhs"
 	"hakurei.app/container/seccomp"
 	"hakurei.app/container/std"
 	"hakurei.app/container/vfs"
@@ -29,6 +30,45 @@ import (
 	"hakurei.app/message"
 )
 
+// Note: this package requires cgo, which is unavailable in the Go playground.
+func Example() {
+	// Must be called early if the current process starts containers.
+	container.TryArgv0(nil)
+
+	// Configure the container.
+	z := container.New(context.Background(), nil)
+	z.Hostname = "hakurei-example"
+	z.Proc(fhs.AbsProc).Dev(fhs.AbsDev, true)
+	z.Stdin, z.Stdout, z.Stderr = os.Stdin, os.Stdout, os.Stderr
+
+	// Bind / for demonstration.
+	z.Bind(fhs.AbsRoot, fhs.AbsRoot, 0)
+	if name, err := exec.LookPath("hostname"); err != nil {
+		panic(err)
+	} else {
+		z.Path = check.MustAbs(name)
+	}
+
+	// This completes the first stage of container setup and starts the container init process.
+	// The new process blocks until the Serve method is called.
+	if err := z.Start(); err != nil {
+		panic(err)
+	}
+
+	// This serves the setup payload to the container init process,
+	// starting the second stage of container setup.
+	if err := z.Serve(); err != nil {
+		panic(err)
+	}
+
+	// Must be called if the Start method succeeds.
+	if err := z.Wait(); err != nil {
+		panic(err)
+	}
+
+	// Output: hakurei-example
+}
+
 func TestStartError(t *testing.T) {
 	t.Parallel()
 
@@ -722,12 +762,14 @@ func TestMain(m *testing.M) {
 func helperNewContainerLibPaths(ctx context.Context, libPaths *[]*check.Absolute, args ...string) (c *container.Container) {
 	msg := message.New(nil)
 	msg.SwapVerbose(testing.Verbose())
+	executable := check.MustAbs(container.MustExecutable(msg))
+
 	c = container.NewCommand(ctx, msg, absHelperInnerPath, "helper", args...)
 	c.Env = append(c.Env, envDoCheck+"=1")
-	c.Bind(check.MustAbs(os.Args[0]), absHelperInnerPath, 0)
+	c.Bind(executable, absHelperInnerPath, 0)
 
 	// in case test has cgo enabled
-	if entries, err := ldd.Exec(ctx, msg, os.Args[0]); err != nil {
+	if entries, err := ldd.Resolve(ctx, msg, executable); err != nil {
 		log.Fatalf("ldd: %v", err)
 	} else {
 		*libPaths = ldd.Path(entries)

container/dispatcher_test.go

@@ -162,7 +162,8 @@ func checkSimple(t *testing.T, fname string, testCases []simpleTestCase) {
 			t.Parallel()
 
 			wait4signal := make(chan struct{})
-			k := &kstub{wait4signal, stub.New(t, func(s *stub.Stub[syscallDispatcher]) syscallDispatcher { return &kstub{wait4signal, s} }, tc.want)}
+			lockNotify := make(chan struct{})
+			k := &kstub{wait4signal, lockNotify, stub.New(t, func(s *stub.Stub[syscallDispatcher]) syscallDispatcher { return &kstub{wait4signal, lockNotify, s} }, tc.want)}
 			defer stub.HandleExit(t)
 			if err := tc.f(k); !reflect.DeepEqual(err, tc.wantErr) {
 				t.Errorf("%s: error = %v, want %v", fname, err, tc.wantErr)
@@ -200,8 +201,8 @@ func checkOpBehaviour(t *testing.T, testCases []opBehaviourTestCase) {
 				t.Helper()
 				t.Parallel()
 
-				k := &kstub{nil, stub.New(t,
-					func(s *stub.Stub[syscallDispatcher]) syscallDispatcher { return &kstub{nil, s} },
+				k := &kstub{nil, nil, stub.New(t,
+					func(s *stub.Stub[syscallDispatcher]) syscallDispatcher { return &kstub{nil, nil, s} },
 					stub.Expect{Calls: slices.Concat(tc.early, []stub.Call{{Name: stub.CallSeparator}}, tc.apply)},
 				)}
 				state := &setupState{Params: tc.params, Msg: k}
@@ -322,12 +323,19 @@ const (
 
 type kstub struct {
 	wait4signal chan struct{}
+	lockNotify  chan struct{}
 	*stub.Stub[syscallDispatcher]
 }
 
 func (k *kstub) new(f func(k syscallDispatcher)) { k.Helper(); k.New(f) }
 
-func (k *kstub) lockOSThread() { k.Helper(); k.Expects("lockOSThread") }
+func (k *kstub) lockOSThread() {
+	k.Helper()
+	expect := k.Expects("lockOSThread")
+	if k.lockNotify != nil && expect.Ret == magicWait4Signal {
+		<-k.lockNotify
+	}
+}
 
 func (k *kstub) setPtracer(pid uintptr) error {
 	k.Helper()
@@ -472,6 +480,10 @@ func (k *kstub) notify(c chan<- os.Signal, sig ...os.Signal) {
 		k.FailNow()
 	}
 
+	if k.lockNotify != nil && expect.Ret == magicWait4Signal {
+		defer close(k.lockNotify)
+	}
+
 	// export channel for external instrumentation
 	if chanf, ok := expect.Args[0].(func(c chan<- os.Signal)); ok && chanf != nil {
 		chanf(c)
@@ -759,7 +771,8 @@ func (k *kstub) checkMsg(msg message.Msg) {
 }
 
 func (k *kstub) GetLogger() *log.Logger { panic("unreachable") }
-func (k *kstub) IsVerbose() bool        { panic("unreachable") }
+
+func (k *kstub) IsVerbose() bool { k.Helper(); return k.Expects("isVerbose").Ret.(bool) }
 
 func (k *kstub) SwapVerbose(verbose bool) bool {
 	k.Helper()

container/errors.go

@@ -7,31 +7,36 @@ import (
 
 	"hakurei.app/container/check"
 	"hakurei.app/container/vfs"
+	"hakurei.app/message"
 )
 
 // messageFromError returns a printable error message for a supported concrete type.
-func messageFromError(err error) (string, bool) {
-	if m, ok := messagePrefixP[MountError]("cannot ", err); ok {
-		return m, ok
+func messageFromError(err error) (m string, ok bool) {
+	if m, ok = messagePrefixP[MountError]("cannot ", err); ok {
+		return
 	}
-	if m, ok := messagePrefixP[os.PathError]("cannot ", err); ok {
-		return m, ok
+	if m, ok = messagePrefixP[os.PathError]("cannot ", err); ok {
+		return
 	}
-	if m, ok := messagePrefixP[check.AbsoluteError]("", err); ok {
-		return m, ok
+	if m, ok = messagePrefixP[check.AbsoluteError](zeroString, err); ok {
+		return
 	}
-	if m, ok := messagePrefix[OpRepeatError]("", err); ok {
-		return m, ok
+	if m, ok = messagePrefix[OpRepeatError](zeroString, err); ok {
+		return
 	}
-	if m, ok := messagePrefix[OpStateError]("", err); ok {
-		return m, ok
+	if m, ok = messagePrefix[OpStateError](zeroString, err); ok {
+		return
 	}
 
-	if m, ok := messagePrefixP[vfs.DecoderError]("cannot ", err); ok {
-		return m, ok
+	if m, ok = messagePrefixP[vfs.DecoderError]("cannot ", err); ok {
+		return
 	}
-	if m, ok := messagePrefix[TmpfsSizeError]("", err); ok {
-		return m, ok
+	if m, ok = messagePrefix[TmpfsSizeError](zeroString, err); ok {
+		return
+	}
+
+	if m, ok = message.GetMessage(err); ok {
+		return
 	}
 
 	return zeroString, false

container/fhs/abs.go

@@ -8,8 +8,10 @@ import (
 
 /* constants in this file bypass abs check, be extremely careful when changing them! */
 
+// unsafeAbs returns check.Absolute on any string value.
+//
 //go:linkname unsafeAbs hakurei.app/container/check.unsafeAbs
-func unsafeAbs(_ string) *check.Absolute
+func unsafeAbs(pathname string) *check.Absolute
 
 var (
 	// AbsRoot is [Root] as [check.Absolute].
@@ -34,6 +36,8 @@ var (
 
 	// AbsDev is [Dev] as [check.Absolute].
 	AbsDev = unsafeAbs(Dev)
+	// AbsDevShm is [DevShm] as [check.Absolute].
+	AbsDevShm = unsafeAbs(DevShm)
 	// AbsProc is [Proc] as [check.Absolute].
 	AbsProc = unsafeAbs(Proc)
 	// AbsSys is [Sys] as [check.Absolute].

container/fhs/fhs.go

@@ -29,6 +29,8 @@ const (
 
 	// Dev points to the root directory for device nodes.
 	Dev = "/dev/"
+	// DevShm is the place for POSIX shared memory segments, as created via shm_open(3).
+	DevShm = "/dev/shm/"
 	// Proc points to a virtual kernel file system exposing the process list and other functionality.
 	Proc = "/proc/"
 	// ProcSys points to a hierarchy below /proc/ that exposes a number of kernel tunables.

container/init.go

@@ -1,6 +1,7 @@
 package container
 
 import (
+	"context"
 	"errors"
 	"fmt"
 	"log"
@@ -9,6 +10,8 @@ import (
 	"path"
 	"slices"
 	"strconv"
+	"sync"
+	"sync/atomic"
 	. "syscall"
 	"time"
 
@@ -18,24 +21,28 @@ import (
 )
 
 const (
-	/* intermediate tmpfs mount point
+	/* intermediateHostPath is the pathname of the intermediate tmpfs mount point.
 
-	this path might seem like a weird choice, however there are many good reasons to use it:
-	- the contents of this path is never exposed to the container:
-	  the tmpfs root established here effectively becomes anonymous after pivot_root
-	- it is safe to assume this path exists and is a directory:
-	  this program will not work correctly without a proper /proc and neither will most others
-	- this path belongs to the container init:
-	  the container init is not any more privileged or trusted than the rest of the container
-	- this path is only accessible by init and root:
-	  the container init sets SUID_DUMP_DISABLE and terminates if that fails;
+	This path might seem like a weird choice, however there are many good reasons to use it:
+	- The contents of this path is never exposed to the container:
+	  The tmpfs root established here effectively becomes anonymous after pivot_root.
+	- It is safe to assume this path exists and is a directory:
+	  This program will not work correctly without a proper /proc and neither will most others.
+	- This path belongs to the container init:
+	  The container init is not any more privileged or trusted than the rest of the container.
+	- This path is only accessible by init and root:
+	  The container init sets SUID_DUMP_DISABLE and terminates if that fails.
 
-	it should be noted that none of this should become relevant at any point since the resulting
-	intermediate root tmpfs should be effectively anonymous */
+	It should be noted that none of this should become relevant at any point since the resulting
+	intermediate root tmpfs should be effectively anonymous. */
 	intermediateHostPath = fhs.Proc + "self/fd"
 
-	// setup params file descriptor
+	// setupEnv is the name of the environment variable holding the string representation of
+	// the read end file descriptor of the setup params pipe.
 	setupEnv = "HAKUREI_SETUP"
+
+	// exitUnexpectedWait4 is the exit code if wait4 returns an unexpected errno.
+	exitUnexpectedWait4 = 2
 )
 
 type (
@@ -49,6 +56,8 @@ type (
 		early(state *setupState, k syscallDispatcher) error
 		// apply is called in intermediate root.
 		apply(state *setupState, k syscallDispatcher) error
+		// late is called right before starting the initial process.
+		late(state *setupState, k syscallDispatcher) error
 
 		// prefix returns a log message prefix, and whether this Op prints no identifying message on its own.
 		prefix() (string, bool)
@@ -61,11 +70,29 @@ type (
 	// setupState persists context between Ops.
 	setupState struct {
 		nonrepeatable uintptr
+
+		// Whether early reaping has concluded. Must only be accessed in the wait4 loop.
+		processConcluded bool
+		// Process to syscall.WaitStatus populated in the wait4 loop. Freed after early reaping concludes.
+		process map[int]WaitStatus
+		// Synchronises access to process.
+		processMu sync.RWMutex
+
 		*Params
+		context.Context
 		message.Msg
 	}
 )
 
+// terminated returns whether the specified pid has been reaped, and its
+// syscall.WaitStatus if it had. This is only usable by [Op].
+func (state *setupState) terminated(pid int) (wstatus WaitStatus, ok bool) {
+	state.processMu.RLock()
+	wstatus, ok = state.process[pid]
+	state.processMu.RUnlock()
+	return
+}
+
 // Grow grows the slice Ops points to using [slices.Grow].
 func (f *Ops) Grow(n int) { *f = slices.Grow(*f, n) }
 
@@ -180,7 +207,9 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 		k.fatalf(msg, "cannot make / rslave: %v", optionalErrorUnwrap(err))
 	}
 
-	state := &setupState{Params: &params.Params, Msg: msg}
+	ctx, cancel := context.WithCancel(context.Background())
+	state := &setupState{process: make(map[int]WaitStatus), Params: &params.Params, Msg: msg, Context: ctx}
+	defer cancel()
 
 	/* early is called right before pivot_root into intermediate root;
 	this step is mostly for gathering information that would otherwise be difficult to obtain
@@ -330,6 +359,97 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 	}
 	k.umask(oldmask)
 
+	// winfo represents an exited process from wait4.
+	type winfo struct {
+		wpid    int
+		wstatus WaitStatus
+	}
+
+	// info is closed as the wait4 thread terminates
+	// when there are no longer any processes left to reap
+	info := make(chan winfo, 1)
+
+	// whether initial process has started
+	var initialProcessStarted atomic.Bool
+
+	k.new(func(k syscallDispatcher) {
+		k.lockOSThread()
+
+	wait4:
+		var (
+			err     error
+			wpid    = -2
+			wstatus WaitStatus
+
+			// whether initial process has started
+			started bool
+		)
+
+		// keep going until no child process is left
+		for wpid != -1 {
+			if err != nil {
+				break
+			}
+
+			if wpid != -2 {
+				if !state.processConcluded {
+					state.processMu.Lock()
+					if state.process == nil {
+						// early reaping has already concluded at this point
+						state.processConcluded = true
+						info <- winfo{wpid, wstatus}
+					} else {
+						// initial process has not yet been created, and the
+						// info channel is not yet being received from
+						state.process[wpid] = wstatus
+					}
+					state.processMu.Unlock()
+				} else {
+					info <- winfo{wpid, wstatus}
+				}
+			}
+
+			if !started {
+				started = initialProcessStarted.Load()
+			}
+
+			err = EINTR
+			for errors.Is(err, EINTR) {
+				wpid, err = k.wait4(-1, &wstatus, 0, nil)
+			}
+		}
+
+		if !errors.Is(err, ECHILD) {
+			k.printf(msg, "unexpected wait4 response: %v", err)
+		} else if !started {
+			// initial process has not yet been reached and all daemons
+			// terminated or none were started in the first place
+			time.Sleep(500 * time.Microsecond)
+			goto wait4
+		}
+
+		close(info)
+	})
+
+	// called right before startup of initial process, all state changes to the
+	// current process is prohibited during late
+	for i, op := range *params.Ops {
+		// ops already checked during early setup
+		if err := op.late(state, k); err != nil {
+			if m, ok := messageFromError(err); ok {
+				k.fatal(msg, m)
+			} else if errors.Is(err, context.DeadlineExceeded) {
+				k.fatalf(msg, "%s deadline exceeded", op.String())
+			} else {
+				k.fatalf(msg, "cannot complete op at index %d: %v", i, err)
+			}
+		}
+	}
+	// early reaping has concluded, this must happen before initial process is created
+	state.processMu.Lock()
+	state.process = nil
+	state.processMu.Unlock()
+
 	if err := closeSetup(); err != nil {
 		k.fatalf(msg, "cannot close setup pipe: %v", err)
 	}
@@ -341,50 +461,11 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 	cmd.ExtraFiles = extraFiles
 	cmd.Dir = params.Dir.String()
 
-	msg.Verbosef("starting initial program %s", params.Path)
+	msg.Verbosef("starting initial process %s", params.Path)
 	if err := k.start(cmd); err != nil {
 		k.fatalf(msg, "%v", err)
 	}
-
-	type winfo struct {
-		wpid    int
-		wstatus WaitStatus
-	}
-
-	// info is closed as the wait4 thread terminates
-	// when there are no longer any processes left to reap
-	info := make(chan winfo, 1)
-
-	k.new(func(k syscallDispatcher) {
-		k.lockOSThread()
-
-		var (
-			err     error
-			wpid    = -2
-			wstatus WaitStatus
-		)
-
-		// keep going until no child process is left
-		for wpid != -1 {
-			if err != nil {
-				break
-			}
-
-			if wpid != -2 {
-				info <- winfo{wpid, wstatus}
-			}
-
-			err = EINTR
-			for errors.Is(err, EINTR) {
-				wpid, err = k.wait4(-1, &wstatus, 0, nil)
-			}
-		}
-		if !errors.Is(err, ECHILD) {
-			k.printf(msg, "unexpected wait4 response: %v", err)
-		}
-
-		close(info)
-	})
+	initialProcessStarted.Store(true)
 
 	// handle signals to dump withheld messages
 	sig := make(chan os.Signal, 2)
@@ -394,7 +475,7 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 	// closed after residualProcessTimeout has elapsed after initial process death
 	timeout := make(chan struct{})
 
-	r := 2
+	r := exitUnexpectedWait4
 	for {
 		select {
 		case s := <-sig:
@@ -426,6 +507,9 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 			}
 
 			if w.wpid == cmd.Process.Pid {
+				// cancel Op context early
+				cancel()
+
 				// start timeout early
 				go func() { time.Sleep(params.AdoptWaitDelay); close(timeout) }()
 

container/init_test.go

@@ -1983,11 +1983,20 @@ func TestInitEntrypoint(t *testing.T) {
 				call("newFile", stub.ExpectArgs{uintptr(0x3a), "extra file 0"}, (*os.File)(nil), nil),
 				call("newFile", stub.ExpectArgs{uintptr(0x3b), "extra file 1"}, (*os.File)(nil), nil),
 				call("umask", stub.ExpectArgs{022}, 0, nil),
+				call("New", stub.ExpectArgs{}, nil, nil),
 				call("fatalf", stub.ExpectArgs{"cannot close setup pipe: %v", []any{stub.UniqueError(13)}}, nil, nil),
 				call("verbosef", stub.ExpectArgs{"starting initial program %s", []any{check.MustAbs("/run/current-system/sw/bin/bash")}}, nil, nil),
 				call("start", stub.ExpectArgs{"/run/current-system/sw/bin/bash", []string{"bash", "-c", "false"}, ([]string)(nil), "/.hakurei/nonexistent"}, nil, stub.UniqueError(12)),
 				call("fatalf", stub.ExpectArgs{"%v", []any{stub.UniqueError(12)}}, nil, nil),
 			},
+
+			/* wait4 */
+			Tracks: []stub.Expect{{Calls: []stub.Call{
+				call("lockOSThread", stub.ExpectArgs{}, magicWait4Signal, nil),
+
+				// this terminates the goroutine at the call, preventing it from leaking while preserving behaviour
+				call("wait4", stub.ExpectArgs{-1, nil, 0, nil, stub.PanicExit}, 0, syscall.ECHILD),
+			}}},
 		}, nil},
 
 		{"lowlastcap signaled cancel forward error", func(k *kstub) error { initEntrypoint(k, k); return nil }, stub.Expect{
@@ -2062,11 +2071,11 @@ func TestInitEntrypoint(t *testing.T) {
 				call("newFile", stub.ExpectArgs{uintptr(0x3a), "extra file 0"}, (*os.File)(nil), nil),
 				call("newFile", stub.ExpectArgs{uintptr(0x3b), "extra file 1"}, (*os.File)(nil), nil),
 				call("umask", stub.ExpectArgs{022}, 0, nil),
+				call("New", stub.ExpectArgs{}, nil, nil),
 				call("fatalf", stub.ExpectArgs{"cannot close setup pipe: %v", []any{stub.UniqueError(10)}}, nil, nil),
 				call("verbosef", stub.ExpectArgs{"starting initial program %s", []any{check.MustAbs("/run/current-system/sw/bin/bash")}}, nil, nil),
 				call("start", stub.ExpectArgs{"/run/current-system/sw/bin/bash", []string{"bash", "-c", "false"}, ([]string)(nil), "/.hakurei/nonexistent"}, &os.Process{Pid: 0xbad}, nil),
-				call("New", stub.ExpectArgs{}, nil, nil),
-				call("notify", stub.ExpectArgs{func(c chan<- os.Signal) { c <- CancelSignal }, []os.Signal{CancelSignal, os.Interrupt, syscall.SIGTERM, syscall.SIGQUIT}}, nil, nil),
+				call("notify", stub.ExpectArgs{func(c chan<- os.Signal) { c <- CancelSignal }, []os.Signal{CancelSignal, os.Interrupt, syscall.SIGTERM, syscall.SIGQUIT}}, magicWait4Signal, nil),
 				call("verbose", stub.ExpectArgs{[]any{"forwarding context cancellation"}}, nil, nil),
 				// magicWait4Signal as ret causes wait4 stub to unblock
 				call("signal", stub.ExpectArgs{"/run/current-system/sw/bin/bash", []string{"bash", "-c", "false"}, ([]string)(nil), "/.hakurei/nonexistent", os.Interrupt}, magicWait4Signal, stub.UniqueError(9)),
@@ -2081,7 +2090,7 @@ func TestInitEntrypoint(t *testing.T) {
 
 			/* wait4 */
 			Tracks: []stub.Expect{{Calls: []stub.Call{
-				call("lockOSThread", stub.ExpectArgs{}, nil, nil),
+				call("lockOSThread", stub.ExpectArgs{}, magicWait4Signal, nil),
 
 				// magicWait4Signal as args[4] causes this to block until simulated signal is delivered
 				call("wait4", stub.ExpectArgs{-1, syscall.WaitStatus(0xfade01ce), 0, nil, magicWait4Signal}, 0xbad, nil),
@@ -2162,11 +2171,11 @@ func TestInitEntrypoint(t *testing.T) {
 				call("newFile", stub.ExpectArgs{uintptr(0x3a), "extra file 0"}, (*os.File)(nil), nil),
 				call("newFile", stub.ExpectArgs{uintptr(0x3b), "extra file 1"}, (*os.File)(nil), nil),
 				call("umask", stub.ExpectArgs{022}, 0, nil),
+				call("New", stub.ExpectArgs{}, nil, nil),
 				call("fatalf", stub.ExpectArgs{"cannot close setup pipe: %v", []any{stub.UniqueError(7)}}, nil, nil),
 				call("verbosef", stub.ExpectArgs{"starting initial program %s", []any{check.MustAbs("/run/current-system/sw/bin/bash")}}, nil, nil),
 				call("start", stub.ExpectArgs{"/run/current-system/sw/bin/bash", []string{"bash", "-c", "false"}, ([]string)(nil), "/.hakurei/nonexistent"}, &os.Process{Pid: 0xbad}, nil),
-				call("New", stub.ExpectArgs{}, nil, nil),
-				call("notify", stub.ExpectArgs{func(c chan<- os.Signal) { c <- syscall.SIGQUIT }, []os.Signal{CancelSignal, os.Interrupt, syscall.SIGTERM, syscall.SIGQUIT}}, nil, nil),
+				call("notify", stub.ExpectArgs{func(c chan<- os.Signal) { c <- syscall.SIGQUIT }, []os.Signal{CancelSignal, os.Interrupt, syscall.SIGTERM, syscall.SIGQUIT}}, magicWait4Signal, nil),
 				call("verbosef", stub.ExpectArgs{"got %s, forwarding to initial process", []any{"quit"}}, nil, nil),
 				// magicWait4Signal as ret causes wait4 stub to unblock
 				call("signal", stub.ExpectArgs{"/run/current-system/sw/bin/bash", []string{"bash", "-c", "false"}, ([]string)(nil), "/.hakurei/nonexistent", syscall.SIGQUIT}, magicWait4Signal, stub.UniqueError(0xfe)),
@@ -2181,7 +2190,7 @@ func TestInitEntrypoint(t *testing.T) {
 
 			/* wait4 */
 			Tracks: []stub.Expect{{Calls: []stub.Call{
-				call("lockOSThread", stub.ExpectArgs{}, nil, nil),
+				call("lockOSThread", stub.ExpectArgs{}, magicWait4Signal, nil),
 
 				// magicWait4Signal as args[4] causes this to block until simulated signal is delivered
 				call("wait4", stub.ExpectArgs{-1, syscall.WaitStatus(0xfade01ce), 0, nil, magicWait4Signal}, 0xbad, nil),
@@ -2262,11 +2271,11 @@ func TestInitEntrypoint(t *testing.T) {
 				call("newFile", stub.ExpectArgs{uintptr(0x3a), "extra file 0"}, (*os.File)(nil), nil),
 				call("newFile", stub.ExpectArgs{uintptr(0x3b), "extra file 1"}, (*os.File)(nil), nil),
 				call("umask", stub.ExpectArgs{022}, 0, nil),
+				call("New", stub.ExpectArgs{}, nil, nil),
 				call("fatalf", stub.ExpectArgs{"cannot close setup pipe: %v", []any{stub.UniqueError(7)}}, nil, nil),
 				call("verbosef", stub.ExpectArgs{"starting initial program %s", []any{check.MustAbs("/run/current-system/sw/bin/bash")}}, nil, nil),
 				call("start", stub.ExpectArgs{"/run/current-system/sw/bin/bash", []string{"bash", "-c", "false"}, ([]string)(nil), "/.hakurei/nonexistent"}, &os.Process{Pid: 0xbad}, nil),
-				call("New", stub.ExpectArgs{}, nil, nil),
-				call("notify", stub.ExpectArgs{func(c chan<- os.Signal) { c <- os.Interrupt }, []os.Signal{CancelSignal, os.Interrupt, syscall.SIGTERM, syscall.SIGQUIT}}, nil, nil),
+				call("notify", stub.ExpectArgs{func(c chan<- os.Signal) { c <- os.Interrupt }, []os.Signal{CancelSignal, os.Interrupt, syscall.SIGTERM, syscall.SIGQUIT}}, magicWait4Signal, nil),
 				call("verbosef", stub.ExpectArgs{"got %s", []any{"interrupt"}}, nil, nil),
 				call("beforeExit", stub.ExpectArgs{}, nil, nil),
 				call("exit", stub.ExpectArgs{0}, nil, nil),
@@ -2274,7 +2283,7 @@ func TestInitEntrypoint(t *testing.T) {
 
 			/* wait4 */
 			Tracks: []stub.Expect{{Calls: []stub.Call{
-				call("lockOSThread", stub.ExpectArgs{}, nil, nil),
+				call("lockOSThread", stub.ExpectArgs{}, magicWait4Signal, nil),
 
 				// this terminates the goroutine at the call, preventing it from leaking while preserving behaviour
 				call("wait4", stub.ExpectArgs{-1, nil, 0, nil, stub.PanicExit}, 0, syscall.ECHILD),
@@ -2353,11 +2362,11 @@ func TestInitEntrypoint(t *testing.T) {
 				call("newFile", stub.ExpectArgs{uintptr(0x3a), "extra file 0"}, (*os.File)(nil), nil),
 				call("newFile", stub.ExpectArgs{uintptr(0x3b), "extra file 1"}, (*os.File)(nil), nil),
 				call("umask", stub.ExpectArgs{022}, 0, nil),
+				call("New", stub.ExpectArgs{}, nil, nil),
 				call("fatalf", stub.ExpectArgs{"cannot close setup pipe: %v", []any{stub.UniqueError(5)}}, nil, nil),
 				call("verbosef", stub.ExpectArgs{"starting initial program %s", []any{check.MustAbs("/run/current-system/sw/bin/bash")}}, nil, nil),
 				call("start", stub.ExpectArgs{"/run/current-system/sw/bin/bash", []string{"bash", "-c", "false"}, ([]string)(nil), "/.hakurei/nonexistent"}, &os.Process{Pid: 0xbad}, nil),
-				call("New", stub.ExpectArgs{}, nil, nil),
-				call("notify", stub.ExpectArgs{nil, []os.Signal{CancelSignal, os.Interrupt, syscall.SIGTERM, syscall.SIGQUIT}}, nil, nil),
+				call("notify", stub.ExpectArgs{nil, []os.Signal{CancelSignal, os.Interrupt, syscall.SIGTERM, syscall.SIGQUIT}}, magicWait4Signal, nil),
 				call("verbose", stub.ExpectArgs{[]any{os.ErrInvalid.Error()}}, nil, nil),
 				call("verbose", stub.ExpectArgs{[]any{os.ErrInvalid.Error()}}, nil, nil),
 				call("verbosef", stub.ExpectArgs{"initial process exited with signal %s", []any{syscall.Signal(0x4e)}}, nil, nil),
@@ -2368,7 +2377,7 @@ func TestInitEntrypoint(t *testing.T) {
 
 			/* wait4 */
 			Tracks: []stub.Expect{{Calls: []stub.Call{
-				call("lockOSThread", stub.ExpectArgs{}, nil, nil),
+				call("lockOSThread", stub.ExpectArgs{}, magicWait4Signal, nil),
 
 				call("wait4", stub.ExpectArgs{-1, syscall.WaitStatus(0xfade01ce), 0, nil}, 0xbad, nil),
 				// this terminates the goroutine at the call, preventing it from leaking while preserving behaviour
@@ -2448,11 +2457,11 @@ func TestInitEntrypoint(t *testing.T) {
 				call("newFile", stub.ExpectArgs{uintptr(0x3a), "extra file 0"}, (*os.File)(nil), nil),
 				call("newFile", stub.ExpectArgs{uintptr(0x3b), "extra file 1"}, (*os.File)(nil), nil),
 				call("umask", stub.ExpectArgs{022}, 0, nil),
+				call("New", stub.ExpectArgs{}, nil, nil),
 				call("fatalf", stub.ExpectArgs{"cannot close setup pipe: %v", []any{stub.UniqueError(3)}}, nil, nil),
 				call("verbosef", stub.ExpectArgs{"starting initial program %s", []any{check.MustAbs("/run/current-system/sw/bin/bash")}}, nil, nil),
 				call("start", stub.ExpectArgs{"/run/current-system/sw/bin/bash", []string{"bash", "-c", "false"}, ([]string)(nil), "/.hakurei/nonexistent"}, &os.Process{Pid: 0xbad}, nil),
-				call("New", stub.ExpectArgs{}, nil, nil),
-				call("notify", stub.ExpectArgs{nil, []os.Signal{CancelSignal, os.Interrupt, syscall.SIGTERM, syscall.SIGQUIT}}, nil, nil),
+				call("notify", stub.ExpectArgs{nil, []os.Signal{CancelSignal, os.Interrupt, syscall.SIGTERM, syscall.SIGQUIT}}, magicWait4Signal, nil),
 				call("verbose", stub.ExpectArgs{[]any{os.ErrInvalid.Error()}}, nil, nil),
 				call("verbose", stub.ExpectArgs{[]any{os.ErrInvalid.Error()}}, nil, nil),
 				call("verbosef", stub.ExpectArgs{"initial process exited with signal %s", []any{syscall.Signal(0x4e)}}, nil, nil),
@@ -2462,7 +2471,7 @@ func TestInitEntrypoint(t *testing.T) {
 
 			/* wait4 */
 			Tracks: []stub.Expect{{Calls: []stub.Call{
-				call("lockOSThread", stub.ExpectArgs{}, nil, nil),
+				call("lockOSThread", stub.ExpectArgs{}, magicWait4Signal, nil),
 
 				call("wait4", stub.ExpectArgs{-1, nil, 0, nil}, 0, syscall.EINTR),
 				call("wait4", stub.ExpectArgs{-1, nil, 0, nil}, 0, syscall.EINTR),
@@ -2586,11 +2595,11 @@ func TestInitEntrypoint(t *testing.T) {
 				call("newFile", stub.ExpectArgs{uintptr(0x3a), "extra file 0"}, (*os.File)(nil), nil),
 				call("newFile", stub.ExpectArgs{uintptr(0x3b), "extra file 1"}, (*os.File)(nil), nil),
 				call("umask", stub.ExpectArgs{022}, 0, nil),
+				call("New", stub.ExpectArgs{}, nil, nil),
 				call("fatalf", stub.ExpectArgs{"cannot close setup pipe: %v", []any{stub.UniqueError(1)}}, nil, nil),
 				call("verbosef", stub.ExpectArgs{"starting initial program %s", []any{check.MustAbs("/run/current-system/sw/bin/bash")}}, nil, nil),
 				call("start", stub.ExpectArgs{"/run/current-system/sw/bin/bash", []string{"bash", "-c", "false"}, ([]string)(nil), "/.hakurei/nonexistent"}, &os.Process{Pid: 0xbad}, nil),
-				call("New", stub.ExpectArgs{}, nil, nil),
-				call("notify", stub.ExpectArgs{nil, []os.Signal{CancelSignal, os.Interrupt, syscall.SIGTERM, syscall.SIGQUIT}}, nil, nil),
+				call("notify", stub.ExpectArgs{nil, []os.Signal{CancelSignal, os.Interrupt, syscall.SIGTERM, syscall.SIGQUIT}}, magicWait4Signal, nil),
 				call("verbose", stub.ExpectArgs{[]any{os.ErrInvalid.Error()}}, nil, nil),
 				call("verbose", stub.ExpectArgs{[]any{os.ErrInvalid.Error()}}, nil, nil),
 				call("verbosef", stub.ExpectArgs{"initial process exited with code %d", []any{1}}, nil, nil),
@@ -2600,7 +2609,7 @@ func TestInitEntrypoint(t *testing.T) {
 
 			/* wait4 */
 			Tracks: []stub.Expect{{Calls: []stub.Call{
-				call("lockOSThread", stub.ExpectArgs{}, nil, nil),
+				call("lockOSThread", stub.ExpectArgs{}, magicWait4Signal, nil),
 
 				call("wait4", stub.ExpectArgs{-1, nil, 0, nil}, 0, syscall.EINTR),
 				call("wait4", stub.ExpectArgs{-1, nil, 0, nil}, 0, syscall.EINTR),
@@ -2728,11 +2737,11 @@ func TestInitEntrypoint(t *testing.T) {
 				call("newFile", stub.ExpectArgs{uintptr(11), "extra file 1"}, (*os.File)(nil), nil),
 				call("newFile", stub.ExpectArgs{uintptr(12), "extra file 2"}, (*os.File)(nil), nil),
 				call("umask", stub.ExpectArgs{022}, 0, nil),
+				call("New", stub.ExpectArgs{}, nil, nil),
 				call("fatalf", stub.ExpectArgs{"cannot close setup pipe: %v", []any{stub.UniqueError(0)}}, nil, nil),
 				call("verbosef", stub.ExpectArgs{"starting initial program %s", []any{check.MustAbs("/bin/zsh")}}, nil, nil),
 				call("start", stub.ExpectArgs{"/bin/zsh", []string{"zsh", "-c", "exec vim"}, []string{"DISPLAY=:0"}, "/.hakurei"}, &os.Process{Pid: 0xcafe}, nil),
-				call("New", stub.ExpectArgs{}, nil, nil),
-				call("notify", stub.ExpectArgs{nil, []os.Signal{CancelSignal, os.Interrupt, syscall.SIGTERM, syscall.SIGQUIT}}, nil, nil),
+				call("notify", stub.ExpectArgs{nil, []os.Signal{CancelSignal, os.Interrupt, syscall.SIGTERM, syscall.SIGQUIT}}, magicWait4Signal, nil),
 				call("verbose", stub.ExpectArgs{[]any{os.ErrInvalid.Error()}}, nil, nil),
 				call("verbose", stub.ExpectArgs{[]any{os.ErrInvalid.Error()}}, nil, nil),
 				call("verbose", stub.ExpectArgs{[]any{os.ErrInvalid.Error()}}, nil, nil),
@@ -2743,7 +2752,7 @@ func TestInitEntrypoint(t *testing.T) {
 
 			/* wait4 */
 			Tracks: []stub.Expect{{Calls: []stub.Call{
-				call("lockOSThread", stub.ExpectArgs{}, nil, nil),
+				call("lockOSThread", stub.ExpectArgs{}, magicWait4Signal, nil),
 
 				call("wait4", stub.ExpectArgs{-1, nil, 0, nil}, 0, syscall.EINTR),
 				call("wait4", stub.ExpectArgs{-1, nil, 0, nil}, 0, syscall.EINTR),

container/initbind.go

@@ -90,6 +90,7 @@ func (b *BindMountOp) apply(state *setupState, k syscallDispatcher) error {
 	}
 	return k.bindMount(state, source, target, flags)
 }
+func (b *BindMountOp) late(*setupState, syscallDispatcher) error { return nil }
 
 func (b *BindMountOp) Is(op Op) bool {
 	vb, ok := op.(*BindMountOp)

container/initdaemon.go

@@ -0,0 +1,134 @@
+package container
+
+import (
+	"context"
+	"encoding/gob"
+	"errors"
+	"fmt"
+	"os"
+	"os/exec"
+	"slices"
+	"strconv"
+	"syscall"
+	"time"
+
+	"hakurei.app/container/check"
+	"hakurei.app/container/fhs"
+)
+
+func init() { gob.Register(new(DaemonOp)) }
+
+const (
+	// daemonTimeout is the duration a [DaemonOp] is allowed to block before the
+	// [DaemonOp.Target] marker becomes available.
+	daemonTimeout = 5 * time.Second
+)
+
+// Daemon appends an [Op] that starts a daemon in the container and blocks until
+// [DaemonOp.Target] appears.
+func (f *Ops) Daemon(target, path *check.Absolute, args ...string) *Ops {
+	*f = append(*f, &DaemonOp{target, path, args})
+	return f
+}
+
+// DaemonOp starts a daemon in the container and blocks until Target appears.
+type DaemonOp struct {
+	// Pathname indicating readiness of daemon.
+	Target *check.Absolute
+	// Absolute pathname passed to [exec.Cmd].
+	Path *check.Absolute
+	// Arguments (excl. first) passed to [exec.Cmd].
+	Args []string
+}
+
+// earlyTerminationError is returned by [DaemonOp] when a daemon terminates
+// before [DaemonOp.Target] appears.
+type earlyTerminationError struct {
+	// Returned by [DaemonOp.String].
+	op string
+	// Copied from wait4 loop.
+	wstatus syscall.WaitStatus
+}
+
+func (e *earlyTerminationError) Error() string {
+	res := ""
+	switch {
+	case e.wstatus.Exited():
+		res = "exit status " + strconv.Itoa(e.wstatus.ExitStatus())
+	case e.wstatus.Signaled():
+		res = "signal: " + e.wstatus.Signal().String()
+	case e.wstatus.Stopped():
+		res = "stop signal: " + e.wstatus.StopSignal().String()
+		if e.wstatus.StopSignal() == syscall.SIGTRAP && e.wstatus.TrapCause() != 0 {
+			res += " (trap " + strconv.Itoa(e.wstatus.TrapCause()) + ")"
+		}
+	case e.wstatus.Continued():
+		res = "continued"
+	}
+	if e.wstatus.CoreDump() {
+		res += " (core dumped)"
+	}
+	return res
+}
+
+func (e *earlyTerminationError) Message() string { return e.op + " " + e.Error() }
+
+func (d *DaemonOp) Valid() bool                                { return d != nil && d.Target != nil && d.Path != nil }
+func (d *DaemonOp) early(*setupState, syscallDispatcher) error { return nil }
+func (d *DaemonOp) apply(*setupState, syscallDispatcher) error { return nil }
+func (d *DaemonOp) late(state *setupState, k syscallDispatcher) error {
+	cmd := exec.CommandContext(state.Context, d.Path.String(), d.Args...)
+	cmd.Env = state.Env
+	cmd.Dir = fhs.Root
+	if state.IsVerbose() {
+		cmd.Stdout, cmd.Stderr = os.Stdout, os.Stderr
+	}
+	// WaitDelay: left unset because lifetime is bound by AdoptWaitDelay on cancellation
+	cmd.Cancel = func() error { return cmd.Process.Signal(syscall.SIGTERM) }
+
+	state.Verbosef("starting %s", d.String())
+	if err := k.start(cmd); err != nil {
+		return err
+	}
+
+	deadline := time.Now().Add(daemonTimeout)
+	var wstatusErr error
+
+	for {
+		if _, err := k.stat(d.Target.String()); err != nil {
+			if !errors.Is(err, os.ErrNotExist) {
+				_ = k.signal(cmd, os.Kill)
+				return err
+			}
+
+			if time.Now().After(deadline) {
+				_ = k.signal(cmd, os.Kill)
+				return context.DeadlineExceeded
+			}
+
+			if wstatusErr != nil {
+				return wstatusErr
+			}
+			if wstatus, ok := state.terminated(cmd.Process.Pid); ok {
+				// check once again: process could have satisfied Target between stat and the lookup
+				wstatusErr = &earlyTerminationError{d.String(), wstatus}
+				continue
+			}
+
+			time.Sleep(500 * time.Microsecond)
+			continue
+		}
+
+		state.Verbosef("daemon process %d ready", cmd.Process.Pid)
+		return nil
+	}
+}
+
+func (d *DaemonOp) Is(op Op) bool {
+	vd, ok := op.(*DaemonOp)
+	return ok && d.Valid() && vd.Valid() &&
+		d.Target.Is(vd.Target) && d.Path.Is(vd.Path) &&
+		slices.Equal(d.Args, vd.Args)
+}
+func (*DaemonOp) prefix() (string, bool) { return zeroString, false }
+func (d *DaemonOp) String() string       { return fmt.Sprintf("daemon providing %q", d.Target) }

container/initdaemon_test.go

@@ -0,0 +1,127 @@
+package container
+
+import (
+	"os"
+	"testing"
+
+	"hakurei.app/container/check"
+	"hakurei.app/container/stub"
+	"hakurei.app/message"
+)
+
+func TestEarlyTerminationError(t *testing.T) {
+	t.Parallel()
+
+	testCases := []struct {
+		name string
+		err  error
+		want string
+		msg  string
+	}{
+		{"exited", &earlyTerminationError{
+			`daemon providing "/run/user/1971/pulse/native"`, 127 << 8,
+		}, "exit status 127", `daemon providing "/run/user/1971/pulse/native" exit status 127`},
+	}
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Parallel()
+
+			if got := tc.err.Error(); got != tc.want {
+				t.Errorf("Error: %q, want %q", got, tc.want)
+			}
+			if got := tc.err.(message.Error).Message(); got != tc.msg {
+				t.Errorf("Message: %s, want %s", got, tc.msg)
+			}
+		})
+	}
+}
+
+func TestDaemonOp(t *testing.T) {
+	t.Parallel()
+
+	checkSimple(t, "DaemonOp.late", []simpleTestCase{
+		{"success", func(k *kstub) error {
+			state := setupState{Params: &Params{Env: []string{"\x00"}}, Context: t.Context(), Msg: k}
+			return (&DaemonOp{
+				Target: check.MustAbs("/run/user/1971/pulse/native"),
+				Path:   check.MustAbs("/run/current-system/sw/bin/pipewire-pulse"),
+				Args:   []string{"-v"},
+			}).late(&state, k)
+		}, stub.Expect{Calls: []stub.Call{
+			call("isVerbose", stub.ExpectArgs{}, true, nil),
+			call("verbosef", stub.ExpectArgs{"starting %s", []any{`daemon providing "/run/user/1971/pulse/native"`}}, nil, nil),
+			call("start", stub.ExpectArgs{"/run/current-system/sw/bin/pipewire-pulse", []string{"/run/current-system/sw/bin/pipewire-pulse", "-v"}, []string{"\x00"}, "/"}, &os.Process{Pid: 0xcafe}, nil),
+			call("stat", stub.ExpectArgs{"/run/user/1971/pulse/native"}, isDirFi(false), os.ErrNotExist),
+			call("stat", stub.ExpectArgs{"/run/user/1971/pulse/native"}, isDirFi(false), os.ErrNotExist),
+			call("stat", stub.ExpectArgs{"/run/user/1971/pulse/native"}, isDirFi(false), os.ErrNotExist),
+			call("stat", stub.ExpectArgs{"/run/user/1971/pulse/native"}, isDirFi(false), nil),
+			call("verbosef", stub.ExpectArgs{"daemon process %d ready", []any{0xcafe}}, nil, nil),
+		}}, nil},
+	})
+
+	checkOpsValid(t, []opValidTestCase{
+		{"nil", (*DaemonOp)(nil), false},
+		{"zero", new(DaemonOp), false},
+		{"valid", &DaemonOp{
+			Target: check.MustAbs("/run/user/1971/pulse/native"),
+			Path:   check.MustAbs("/run/current-system/sw/bin/pipewire-pulse"),
+			Args:   []string{"-v"},
+		}, true},
+	})
+
+	checkOpsBuilder(t, []opsBuilderTestCase{
+		{"pipewire-pulse", new(Ops).Daemon(
+			check.MustAbs("/run/user/1971/pulse/native"),
+			check.MustAbs("/run/current-system/sw/bin/pipewire-pulse"), "-v",
+		), Ops{
+			&DaemonOp{
+				Target: check.MustAbs("/run/user/1971/pulse/native"),
+				Path:   check.MustAbs("/run/current-system/sw/bin/pipewire-pulse"),
+				Args:   []string{"-v"},
+			},
+		}},
+	})
+
+	checkOpIs(t, []opIsTestCase{
+		{"zero", new(DaemonOp), new(DaemonOp), false},
+
+		{"args differs", &DaemonOp{
+			Target: check.MustAbs("/run/user/1971/pulse/native"),
+			Path:   check.MustAbs("/run/current-system/sw/bin/pipewire-pulse"),
+			Args:   []string{"-v"},
+		}, &DaemonOp{
+			Target: check.MustAbs("/run/user/1971/pulse/native"),
+			Path:   check.MustAbs("/run/current-system/sw/bin/pipewire-pulse"),
+		}, false},
+
+		{"path differs", &DaemonOp{
+			Target: check.MustAbs("/run/user/1971/pulse/native"),
+			Path:   check.MustAbs("/run/current-system/sw/bin/pipewire"),
+		}, &DaemonOp{
+			Target: check.MustAbs("/run/user/1971/pulse/native"),
+			Path:   check.MustAbs("/run/current-system/sw/bin/pipewire-pulse"),
+		}, false},
+
+		{"target differs", &DaemonOp{
+			Target: check.MustAbs("/run/user/65534/pulse/native"),
+			Path:   check.MustAbs("/run/current-system/sw/bin/pipewire-pulse"),
+		}, &DaemonOp{
+			Target: check.MustAbs("/run/user/1971/pulse/native"),
+			Path:   check.MustAbs("/run/current-system/sw/bin/pipewire-pulse"),
+		}, false},
+
+		{"equals", &DaemonOp{
+			Target: check.MustAbs("/run/user/1971/pulse/native"),
+			Path:   check.MustAbs("/run/current-system/sw/bin/pipewire-pulse"),
+		}, &DaemonOp{
+			Target: check.MustAbs("/run/user/1971/pulse/native"),
+			Path:   check.MustAbs("/run/current-system/sw/bin/pipewire-pulse"),
+		}, true},
+	})
+
+	checkOpMeta(t, []opMetaTestCase{
+		{"pipewire-pulse", &DaemonOp{
+			Target: check.MustAbs("/run/user/1971/pulse/native"),
+		}, zeroString, `daemon providing "/run/user/1971/pulse/native"`},
+	})
+}

container/initdev.go

@@ -126,6 +126,7 @@ func (d *MountDevOp) apply(state *setupState, k syscallDispatcher) error {
 	}
 	return k.mountTmpfs(SourceTmpfs, devShmPath, MS_NOSUID|MS_NODEV, 0, 01777)
 }
+func (d *MountDevOp) late(*setupState, syscallDispatcher) error { return nil }
 
 func (d *MountDevOp) Is(op Op) bool {
 	vd, ok := op.(*MountDevOp)

container/initmkdir.go

@@ -27,6 +27,7 @@ func (m *MkdirOp) early(*setupState, syscallDispatcher) error { return nil }
 func (m *MkdirOp) apply(_ *setupState, k syscallDispatcher) error {
 	return k.mkdirAll(toSysroot(m.Path.String()), m.Perm)
 }
+func (m *MkdirOp) late(*setupState, syscallDispatcher) error { return nil }
 
 func (m *MkdirOp) Is(op Op) bool {
 	vm, ok := op.(*MkdirOp)

container/initoverlay.go

@@ -205,6 +205,8 @@ func (o *MountOverlayOp) apply(state *setupState, k syscallDispatcher) error {
 	return k.mount(SourceOverlay, target, FstypeOverlay, 0, strings.Join(options, check.SpecialOverlayOption))
 }
 
+func (o *MountOverlayOp) late(*setupState, syscallDispatcher) error { return nil }
+
 func (o *MountOverlayOp) Is(op Op) bool {
 	vo, ok := op.(*MountOverlayOp)
 	return ok && o.Valid() && vo.Valid() &&

container/initplace.go

@@ -57,6 +57,7 @@ func (t *TmpfileOp) apply(state *setupState, k syscallDispatcher) error {
 	}
 	return nil
 }
+func (t *TmpfileOp) late(*setupState, syscallDispatcher) error { return nil }
 
 func (t *TmpfileOp) Is(op Op) bool {
 	vt, ok := op.(*TmpfileOp)

container/initproc.go

@@ -28,6 +28,7 @@ func (p *MountProcOp) apply(state *setupState, k syscallDispatcher) error {
 	}
 	return k.mount(SourceProc, target, FstypeProc, MS_NOSUID|MS_NOEXEC|MS_NODEV, zeroString)
 }
+func (p *MountProcOp) late(*setupState, syscallDispatcher) error { return nil }
 
 func (p *MountProcOp) Is(op Op) bool {
 	vp, ok := op.(*MountProcOp)

container/initremount.go

@@ -26,6 +26,7 @@ func (*RemountOp) early(*setupState, syscallDispatcher) error { return nil }
 func (r *RemountOp) apply(state *setupState, k syscallDispatcher) error {
 	return k.remount(state, toSysroot(r.Target.String()), r.Flags)
 }
+func (r *RemountOp) late(*setupState, syscallDispatcher) error { return nil }
 
 func (r *RemountOp) Is(op Op) bool {
 	vr, ok := op.(*RemountOp)

container/initsymlink.go

@@ -50,6 +50,8 @@ func (l *SymlinkOp) apply(state *setupState, k syscallDispatcher) error {
 	return k.symlink(l.LinkName, target)
 }
 
+func (l *SymlinkOp) late(*setupState, syscallDispatcher) error { return nil }
+
 func (l *SymlinkOp) Is(op Op) bool {
 	vl, ok := op.(*SymlinkOp)
 	return ok && l.Valid() && vl.Valid() &&

container/inittmpfs.go

@@ -48,6 +48,7 @@ func (t *MountTmpfsOp) apply(_ *setupState, k syscallDispatcher) error {
 	}
 	return k.mountTmpfs(t.FSName, toSysroot(t.Path.String()), t.Flags, t.Size, t.Perm)
 }
+func (t *MountTmpfsOp) late(*setupState, syscallDispatcher) error { return nil }
 
 func (t *MountTmpfsOp) Is(op Op) bool {
 	vt, ok := op.(*MountTmpfsOp)

container/seccomp/libseccomp-helper.c

@@ -9,136 +9,130 @@
 
 #define LEN(arr) (sizeof(arr) / sizeof((arr)[0]))
 
-int32_t hakurei_scmp_make_filter(int *ret_p, uintptr_t allocate_p,
-                                 uint32_t arch, uint32_t multiarch,
-                                 struct hakurei_syscall_rule *rules,
-                                 size_t rules_sz, hakurei_export_flag flags) {
-  int i;
-  int last_allowed_family;
-  int disallowed;
-  struct hakurei_syscall_rule *rule;
-  void *buf;
-  size_t len = 0;
+int32_t hakurei_scmp_make_filter(
+    int *ret_p, uintptr_t allocate_p,
+    uint32_t arch, uint32_t multiarch,
+    struct hakurei_syscall_rule *rules,
+    size_t rules_sz, hakurei_export_flag flags) {
+    int i;
+    int last_allowed_family;
+    int disallowed;
+    struct hakurei_syscall_rule *rule;
+    void *buf;
+    size_t len = 0;
 
-  int32_t res = 0; /* refer to resPrefix for message */
+    int32_t res = 0; /* refer to resPrefix for message */
 
-  /* Blocklist all but unix, inet, inet6 and netlink */
-  struct {
-    int family;
-    hakurei_export_flag 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, HAKUREI_EXPORT_CAN},
-      {AF_BLUETOOTH, HAKUREI_EXPORT_BLUETOOTH},
-  };
+    /* Blocklist all but unix, inet, inet6 and netlink */
+    struct {
+        int family;
+        hakurei_export_flag 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, HAKUREI_EXPORT_CAN},
+        {AF_BLUETOOTH, HAKUREI_EXPORT_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 (flags & HAKUREI_EXPORT_MULTIARCH && multiarch != 0) {
-      *ret_p = seccomp_arch_add(ctx, multiarch);
-      if (*ret_p < 0 && *ret_p != -EEXIST) {
-        res = 3;
+    scmp_filter_ctx ctx = seccomp_init(SCMP_ACT_ALLOW);
+    if (ctx == NULL) {
+        res = 1;
         goto out;
-      }
-    }
-  }
+    } else
+        errno = 0;
 
-  for (i = 0; i < rules_sz; i++) {
-    rule = &rules[i];
-    assert(rule->m_errno == EPERM || rule->m_errno == ENOSYS);
+    /* 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 (rule->arg)
-      *ret_p = seccomp_rule_add(ctx, SCMP_ACT_ERRNO(rule->m_errno),
-                                rule->syscall, 1, *rule->arg);
-    else
-      *ret_p = seccomp_rule_add(ctx, SCMP_ACT_ERRNO(rule->m_errno),
-                                rule->syscall, 0);
-
-    if (*ret_p == -EFAULT) {
-      res = 4;
-      goto out;
-    } else if (*ret_p < 0) {
-      res = 5;
-      goto out;
-    }
-  }
-
-  /* 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 */
-  last_allowed_family = -1;
-  for (i = 0; i < LEN(socket_family_allowlist); i++) {
-    if (socket_family_allowlist[i].flags_mask != 0 &&
-        (socket_family_allowlist[i].flags_mask & flags) !=
-            socket_family_allowlist[i].flags_mask)
-      continue;
-
-    for (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 (allocate_p == 0) {
-    *ret_p = seccomp_load(ctx);
-    if (*ret_p != 0) {
-      res = 7;
-      goto out;
-    }
-  } else {
-    *ret_p = seccomp_export_bpf_mem(ctx, NULL, &len);
-    if (*ret_p != 0) {
-      res = 6;
-      goto out;
+        if (flags & HAKUREI_EXPORT_MULTIARCH && multiarch != 0) {
+            *ret_p = seccomp_arch_add(ctx, multiarch);
+            if (*ret_p < 0 && *ret_p != -EEXIST) {
+                res = 3;
+                goto out;
+            }
+        }
     }
 
-    buf = hakurei_scmp_allocate(allocate_p, len);
-    if (buf == NULL) {
-      res = 4;
-      goto out;
+    for (i = 0; i < rules_sz; i++) {
+        rule = &rules[i];
+        assert(rule->m_errno == EPERM || rule->m_errno == ENOSYS);
+
+        if (rule->arg)
+            *ret_p = seccomp_rule_add(ctx, SCMP_ACT_ERRNO(rule->m_errno), rule->syscall, 1, *rule->arg);
+        else
+            *ret_p = seccomp_rule_add(ctx, SCMP_ACT_ERRNO(rule->m_errno), rule->syscall, 0);
+
+        if (*ret_p == -EFAULT) {
+            res = 4;
+            goto out;
+        } else if (*ret_p < 0) {
+            res = 5;
+            goto out;
+        }
     }
 
-    *ret_p = seccomp_export_bpf_mem(ctx, buf, &len);
-    if (*ret_p != 0) {
-      res = 6;
-      goto out;
+    /* 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 */
+    last_allowed_family = -1;
+    for (i = 0; i < LEN(socket_family_allowlist); i++) {
+        if (socket_family_allowlist[i].flags_mask != 0 &&
+            (socket_family_allowlist[i].flags_mask & flags) != socket_family_allowlist[i].flags_mask)
+            continue;
+
+        for (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 (allocate_p == 0) {
+        *ret_p = seccomp_load(ctx);
+        if (*ret_p != 0) {
+            res = 7;
+            goto out;
+        }
+    } else {
+        *ret_p = seccomp_export_bpf_mem(ctx, NULL, &len);
+        if (*ret_p != 0) {
+            res = 6;
+            goto out;
+        }
+
+        buf = hakurei_scmp_allocate(allocate_p, len);
+        if (buf == NULL) {
+            res = 4;
+            goto out;
+        }
+
+        *ret_p = seccomp_export_bpf_mem(ctx, buf, &len);
+        if (*ret_p != 0) {
+            res = 6;
+            goto out;
+        }
     }
-  }
 
 out:
-  if (ctx)
-    seccomp_release(ctx);
+    if (ctx)
+        seccomp_release(ctx);
 
-  return res;
+    return res;
 }

container/seccomp/libseccomp-helper.h

@@ -1,25 +1,26 @@
 #include <seccomp.h>
 #include <stdint.h>
 
-#if (SCMP_VER_MAJOR < 2) || (SCMP_VER_MAJOR == 2 && SCMP_VER_MINOR < 5) ||     \
+#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 {
-  HAKUREI_EXPORT_MULTIARCH = 1 << 0,
-  HAKUREI_EXPORT_CAN = 1 << 1,
-  HAKUREI_EXPORT_BLUETOOTH = 1 << 2,
+    HAKUREI_EXPORT_MULTIARCH = 1 << 0,
+    HAKUREI_EXPORT_CAN = 1 << 1,
+    HAKUREI_EXPORT_BLUETOOTH = 1 << 2,
 } hakurei_export_flag;
 
 struct hakurei_syscall_rule {
-  int syscall;
-  int m_errno;
-  struct scmp_arg_cmp *arg;
+    int syscall;
+    int m_errno;
+    struct scmp_arg_cmp *arg;
 };
 
 extern void *hakurei_scmp_allocate(uintptr_t f, size_t len);
-int32_t hakurei_scmp_make_filter(int *ret_p, uintptr_t allocate_p,
-                                 uint32_t arch, uint32_t multiarch,
-                                 struct hakurei_syscall_rule *rules,
-                                 size_t rules_sz, hakurei_export_flag flags);
+int32_t hakurei_scmp_make_filter(
+    int *ret_p, uintptr_t allocate_p,
+    uint32_t arch, uint32_t multiarch,
+    struct hakurei_syscall_rule *rules,
+    size_t rules_sz, hakurei_export_flag flags);

container/stub/exit_test.go

@@ -7,8 +7,10 @@ import (
 	"hakurei.app/container/stub"
 )
 
+// Made available here to check panic recovery behaviour.
+//
 //go:linkname handleExitNew hakurei.app/container/stub.handleExitNew
-func handleExitNew(_ testing.TB)
+func handleExitNew(t testing.TB)
 
 // overrideTFailNow overrides the Fail and FailNow method.
 type overrideTFailNow struct {

dist/comp/_hakurei

@@ -17,7 +17,8 @@ _hakurei_run() {
     '--wayland[Enable connection to Wayland via security-context-v1]' \
     '-X[Enable direct connection to X11]' \
     '--dbus[Enable proxied connection to D-Bus]' \
-    '--pulse[Enable direct connection to PulseAudio]' \
+    '--pipewire[Enable connection to PipeWire via SecurityContext]' \
+    '--pulse[Enable PulseAudio compatibility daemon]' \
     '--dbus-config[Path to session bus proxy config file]: :_files -g "*.json"' \
     '--dbus-system[Path to system bus proxy config file]: :_files -g "*.json"' \
     '--mpris[Allow owning MPRIS D-Bus path]' \

dist/install.sh

@@ -2,7 +2,7 @@
 cd "$(dirname -- "$0")" || exit 1
 
 install -vDm0755 "bin/hakurei" "${HAKUREI_INSTALL_PREFIX}/usr/bin/hakurei"
-install -vDm0755 "bin/hpkg" "${HAKUREI_INSTALL_PREFIX}/usr/bin/hpkg"
+install -vDm0755 "bin/sharefs" "${HAKUREI_INSTALL_PREFIX}/usr/bin/sharefs"
 
 install -vDm4511 "bin/hsu" "${HAKUREI_INSTALL_PREFIX}/usr/bin/hsu"
 if [ ! -f "${HAKUREI_INSTALL_PREFIX}/etc/hsurc" ]; then

dist/release.sh

@@ -10,9 +10,9 @@ cp -rv "dist/comp" "${out}"
 
 go generate ./...
 go build -trimpath -v -o "${out}/bin/" -ldflags "-s -w -buildid= -extldflags '-static'
-  -X hakurei.app/internal.buildVersion=${VERSION}
-  -X hakurei.app/internal.hakureiPath=/usr/bin/hakurei
-  -X hakurei.app/internal.hsuPath=/usr/bin/hsu
+  -X hakurei.app/internal/info.buildVersion=${VERSION}
+  -X hakurei.app/internal/info.hakureiPath=/usr/bin/hakurei
+  -X hakurei.app/internal/info.hsuPath=/usr/bin/hsu
   -X main.hakureiPath=/usr/bin/hakurei" ./...
 
 rm -f "./${out}.tar.gz" && tar -C dist -czf "${out}.tar.gz" "${pname}"

flake.lock

@@ -7,32 +7,32 @@
         ]
       },
       "locked": {
-        "lastModified": 1756679287,
-        "narHash": "sha256-Xd1vOeY9ccDf5VtVK12yM0FS6qqvfUop8UQlxEB+gTQ=",
+        "lastModified": 1765384171,
+        "narHash": "sha256-FuFtkJrW1Z7u+3lhzPRau69E0CNjADku1mLQQflUORo=",
         "owner": "nix-community",
         "repo": "home-manager",
-        "rev": "07fc025fe10487dd80f2ec694f1cd790e752d0e8",
+        "rev": "44777152652bc9eacf8876976fa72cc77ca8b9d8",
         "type": "github"
       },
       "original": {
         "owner": "nix-community",
-        "ref": "release-25.05",
+        "ref": "release-25.11",
         "repo": "home-manager",
         "type": "github"
       }
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1757020766,
-        "narHash": "sha256-PLoSjHRa2bUbi1x9HoXgTx2AiuzNXs54c8omhadyvp0=",
+        "lastModified": 1765311797,
+        "narHash": "sha256-mSD5Ob7a+T2RNjvPvOA1dkJHGVrNVl8ZOrAwBjKBDQo=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "fe83bbdde2ccdc2cb9573aa846abe8363f79a97a",
+        "rev": "09eb77e94fa25202af8f3e81ddc7353d9970ac1b",
         "type": "github"
       },
       "original": {
         "owner": "NixOS",
-        "ref": "nixos-25.05",
+        "ref": "nixos-25.11",
         "repo": "nixpkgs",
         "type": "github"
       }

flake.nix

@@ -2,10 +2,10 @@
   description = "hakurei container tool and nixos module";
 
   inputs = {
-    nixpkgs.url = "github:NixOS/nixpkgs/nixos-25.05";
+    nixpkgs.url = "github:NixOS/nixpkgs/nixos-25.11";
 
     home-manager = {
-      url = "github:nix-community/home-manager/release-25.05";
+      url = "github:nix-community/home-manager/release-25.11";
       inputs.nixpkgs.follows = "nixpkgs";
     };
   };
@@ -69,6 +69,8 @@
             withRace = true;
           };
 
+          sharefs = callPackage ./cmd/sharefs/test { inherit system self; };
+
           hpkg = callPackage ./cmd/hpkg/test { inherit system self; };
 
           formatting = runCommandLocal "check-formatting" { nativeBuildInputs = [ nixfmt-rfc-style ]; } ''
@@ -114,7 +116,7 @@
             inherit (pkgs)
               # passthru.buildInputs
               go
-              gcc
+              clang
 
               # nativeBuildInputs
               pkg-config
@@ -129,9 +131,17 @@
               zstd
               gnutar
               coreutils
+
+              # for check
+              util-linux
+              nettools
               ;
           };
           hsu = pkgs.callPackage ./cmd/hsu/package.nix { inherit (self.packages.${system}) hakurei; };
+          sharefs = pkgs.linkFarm "sharefs" {
+            "bin/sharefs" = "${hakurei}/libexec/sharefs";
+            "bin/mount.fuse.sharefs" = "${hakurei}/libexec/sharefs";
+          };
 
           dist = pkgs.runCommand "${hakurei.name}-dist" { buildInputs = hakurei.targetPkgs ++ [ pkgs.pkgsStatic.musl ]; } ''
             # go requires XDG_CACHE_HOME for the build cache
@@ -144,7 +154,7 @@
                 && chmod -R +w .
 
             export HAKUREI_VERSION="v${hakurei.version}"
-            ./dist/release.sh && mkdir $out && cp -v "dist/hakurei-$HAKUREI_VERSION.tar.gz"* $out
+            CC="clang -O3 -Werror" ./dist/release.sh && mkdir $out && cp -v "dist/hakurei-$HAKUREI_VERSION.tar.gz"* $out
           '';
         }
       );
@@ -156,7 +166,10 @@
           pkgs = nixpkgsFor.${system};
         in
         {
-          default = pkgs.mkShell { buildInputs = hakurei.targetPkgs; };
+          default = pkgs.mkShell {
+            buildInputs = hakurei.targetPkgs;
+            hardeningDisable = [ "fortify" ];
+          };
           withPackage = pkgs.mkShell { buildInputs = [ hakurei ] ++ hakurei.targetPkgs; };
 
           vm =
@@ -181,13 +194,13 @@
                       hakurei =
                         let
                           # this is used for interactive vm testing during development, where tests might be broken
-                          package = self.packages.${pkgs.system}.hakurei.override {
+                          package = self.packages.${pkgs.stdenv.hostPlatform.system}.hakurei.override {
                             buildGoModule = previousArgs: pkgs.pkgsStatic.buildGoModule (previousArgs // { doCheck = false; });
                           };
                         in
                         {
                           inherit package;
-                          hsuPackage = self.packages.${pkgs.system}.hsu.override { hakurei = package; };
+                          hsuPackage = self.packages.${pkgs.stdenv.hostPlatform.system}.hsu.override { hakurei = package; };
                         };
                     };
                   }

helper/deprecated.go

@@ -0,0 +1,73 @@
+// Package helper exposes the internal/helper package.
+//
+// Deprecated: This package will be removed in 0.4.
+package helper
+
+import (
+	"context"
+	"io"
+	"os"
+	"os/exec"
+	"time"
+	_ "unsafe" // for go:linkname
+
+	"hakurei.app/container"
+	"hakurei.app/container/check"
+	"hakurei.app/internal/helper"
+	"hakurei.app/message"
+)
+
+//go:linkname WaitDelay hakurei.app/internal/helper.WaitDelay
+var WaitDelay time.Duration
+
+const (
+	// HakureiHelper is set to 1 when args fd is enabled and 0 otherwise.
+	HakureiHelper = helper.HakureiHelper
+	// HakureiStatus is set to 1 when stat fd is enabled and 0 otherwise.
+	HakureiStatus = helper.HakureiStatus
+)
+
+type Helper = helper.Helper
+
+// NewCheckedArgs returns a checked null-terminated argument writer for a copy of args.
+//
+//go:linkname NewCheckedArgs hakurei.app/internal/helper.NewCheckedArgs
+func NewCheckedArgs(args ...string) (wt io.WriterTo, err error)
+
+// MustNewCheckedArgs returns a checked null-terminated argument writer for a copy of args.
+// If s contains a NUL byte this function panics instead of returning an error.
+//
+//go:linkname MustNewCheckedArgs hakurei.app/internal/helper.MustNewCheckedArgs
+func MustNewCheckedArgs(args ...string) io.WriterTo
+
+// NewDirect initialises a new direct Helper instance with wt as the null-terminated argument writer.
+// Function argF returns an array of arguments passed directly to the child process.
+//
+//go:linkname NewDirect hakurei.app/internal/helper.NewDirect
+func NewDirect(
+	ctx context.Context,
+	name string,
+	wt io.WriterTo,
+	stat bool,
+	argF func(argsFd, statFd int) []string,
+	cmdF func(cmd *exec.Cmd),
+	extraFiles []*os.File,
+) Helper
+
+// New initialises a Helper instance with wt as the null-terminated argument writer.
+//
+//go:linkname New hakurei.app/internal/helper.New
+func New(
+	ctx context.Context,
+	msg message.Msg,
+	pathname *check.Absolute, name string,
+	wt io.WriterTo,
+	stat bool,
+	argF func(argsFd, statFd int) []string,
+	cmdF func(z *container.Container),
+	extraFiles []*os.File,
+) Helper
+
+// InternalHelperStub is an internal function but exported because it is cross-package;
+// it is part of the implementation of the helper stub.
+func InternalHelperStub() { helper.InternalHelperStub() }

helper/proc/deprecated.go

@@ -0,0 +1,63 @@
+// Deprecated: This package will be removed in 0.4.
+package proc
+
+import (
+	"context"
+	"io"
+	"os"
+	"os/exec"
+	"time"
+	_ "unsafe" // for go:linkname
+
+	"hakurei.app/internal/helper/proc"
+)
+
+//go:linkname FulfillmentTimeout hakurei.app/internal/helper/proc.FulfillmentTimeout
+var FulfillmentTimeout time.Duration
+
+// A File is an extra file with deferred initialisation.
+type File = proc.File
+
+// ExtraFilesPre is a linked list storing addresses of [os.File].
+type ExtraFilesPre = proc.ExtraFilesPre
+
+// Fulfill calls the [File.Fulfill] method on all files, starts cmd and blocks until all fulfillment completes.
+//
+//go:linkname Fulfill hakurei.app/internal/helper/proc.Fulfill
+func Fulfill(ctx context.Context,
+	v *[]*os.File, start func() error,
+	files []File, extraFiles *ExtraFilesPre,
+) (err error)
+
+// InitFile initialises f as part of the slice extraFiles points to,
+// and returns its final fd value.
+//
+//go:linkname InitFile hakurei.app/internal/helper/proc.InitFile
+func InitFile(f File, extraFiles *ExtraFilesPre) (fd uintptr)
+
+// BaseFile implements the Init method of the File interface and provides indirect access to extra file state.
+type BaseFile = proc.BaseFile
+
+//go:linkname ExtraFile hakurei.app/internal/helper/proc.ExtraFile
+func ExtraFile(cmd *exec.Cmd, f *os.File) (fd uintptr)
+
+//go:linkname ExtraFileSlice hakurei.app/internal/helper/proc.ExtraFileSlice
+func ExtraFileSlice(extraFiles *[]*os.File, f *os.File) (fd uintptr)
+
+// NewWriterTo returns a [File] that receives content from wt on fulfillment.
+//
+//go:linkname NewWriterTo hakurei.app/internal/helper/proc.NewWriterTo
+func NewWriterTo(wt io.WriterTo) File
+
+// NewStat returns a [File] implementing the behaviour
+// of the receiving end of xdg-dbus-proxy stat fd.
+//
+//go:linkname NewStat hakurei.app/internal/helper/proc.NewStat
+func NewStat(s *io.Closer) File
+
+var (
+	//go:linkname ErrStatFault hakurei.app/internal/helper/proc.ErrStatFault
+	ErrStatFault error
+	//go:linkname ErrStatRead hakurei.app/internal/helper/proc.ErrStatRead
+	ErrStatRead error
+)

hst/config.go

@@ -23,9 +23,54 @@ type Config struct {
 	// System D-Bus proxy configuration.
 	// If set to nil, system bus proxy is disabled.
 	SystemBus *BusConfig `json:"system_bus,omitempty"`
+
 	// Direct access to wayland socket, no attempt is made to attach security-context-v1
 	// and the bare socket is made available to the container.
+	//
+	// This option is unsupported and most likely enables full control over the Wayland
+	// session. Do not set this to true unless you are sure you know what you are doing.
 	DirectWayland bool `json:"direct_wayland,omitempty"`
+	// Direct access to the PipeWire socket established via SecurityContext::Create, no
+	// attempt is made to start the pipewire-pulse server.
+	//
+	// The SecurityContext machinery is fatally flawed, it blindly sets read and execute
+	// bits on all objects for clients with the lowest achievable privilege level (by
+	// setting PW_KEY_ACCESS to "restricted"). This enables them to call any method
+	// targeting any object, and since Registry::Destroy checks for the read and execute bit,
+	// allows the destruction of any object other than PW_ID_CORE as well. This behaviour
+	// is implemented separately in media-session and wireplumber, with the wireplumber
+	// implementation in Lua via an embedded Lua vm. In all known setups, wireplumber is
+	// in use, and there is no known way to change its behaviour and set permissions
+	// differently without replacing the Lua script. Also, since PipeWire relies on these
+	// permissions to work, reducing them is not possible.
+	//
+	// Currently, the only other sandboxed use case is flatpak, which is not aware of
+	// PipeWire and blindly exposes the bare PulseAudio socket to the container (behaves
+	// like DirectPulse). This socket is backed by the pipewire-pulse compatibility daemon,
+	// which obtains client pid via the SO_PEERCRED option. The PipeWire daemon, pipewire-pulse
+	// daemon and the session manager daemon then separately performs the /.flatpak-info hack
+	// described in https://git.gensokyo.uk/security/hakurei/issues/21. Under such use case,
+	// since the client has no direct access to PipeWire, insecure parts of the protocol are
+	// obscured by pipewire-pulse simply not implementing them, and thus hiding the flaws
+	// described above.
+	//
+	// Hakurei does not rely on the /.flatpak-info hack. Instead, a socket is sets up via
+	// SecurityContext. A pipewire-pulse server connected through it achieves the same
+	// permissions as flatpak does via the /.flatpak-info hack and is maintained for the
+	// life of the container.
+	//
+	// This option is unsupported and enables a denial-of-service attack as the sandboxed
+	// client is able to destroy any client object and thus disconnecting them from PipeWire,
+	// or destroy the SecurityContext object preventing any further container creation.
+	// Do not set this to true, it is insecure under any configuration.
+	DirectPipeWire bool `json:"direct_pipewire,omitempty"`
+	// Direct access to PulseAudio socket, no attempt is made to establish pipewire-pulse
+	// server via a PipeWire socket with a SecurityContext attached and the bare socket
+	// is made available to the container.
+	//
+	// This option is unsupported and enables arbitrary code execution as the PulseAudio
+	// server. Do not set this to true, it is insecure under any configuration.
+	DirectPulse bool `json:"direct_pulse,omitempty"`
 
 	// Extra acl updates to perform before setuid.
 	ExtraPerms []ExtraPermConfig `json:"extra_perms,omitempty"`
@@ -49,6 +94,9 @@ var (
 
 	// ErrEnviron is returned by [Config.Validate] if an environment variable name contains '=' or NUL.
 	ErrEnviron = errors.New("invalid environment variable name")
+
+	// ErrInsecure is returned by [Config.Validate] if the configuration is considered insecure.
+	ErrInsecure = errors.New("configuration is insecure")
 )
 
 // Validate checks [Config] and returns [AppError] if an invalid value is encountered.
@@ -95,6 +143,11 @@ func (config *Config) Validate() error {
 		}
 	}
 
+	if et := config.Enablements.Unwrap(); !config.DirectPulse && et&EPulse != 0 {
+		return &AppError{Step: "validate configuration", Err: ErrInsecure,
+			Msg: "enablement PulseAudio is insecure and no longer supported"}
+	}
+
 	return nil
 }
 

hst/config_test.go

@@ -53,6 +53,12 @@ func TestConfigValidate(t *testing.T) {
 			Env:   map[string]string{"TERM\x00": ""},
 		}}, &hst.AppError{Step: "validate configuration", Err: hst.ErrEnviron,
 			Msg: `invalid environment variable "TERM\x00"`}},
+		{"insecure pulse", &hst.Config{Enablements: hst.NewEnablements(hst.EPulse), Container: &hst.ContainerConfig{
+			Home:  fhs.AbsTmp,
+			Shell: fhs.AbsTmp,
+			Path:  fhs.AbsTmp,
+		}}, &hst.AppError{Step: "validate configuration", Err: hst.ErrInsecure,
+			Msg: "enablement PulseAudio is insecure and no longer supported"}},
 		{"valid", &hst.Config{Container: &hst.ContainerConfig{
 			Home:  fhs.AbsTmp,
 			Shell: fhs.AbsTmp,

hst/enablement.go

@@ -17,6 +17,8 @@ const (
 	EX11
 	// EDBus enables the per-container xdg-dbus-proxy daemon.
 	EDBus
+	// EPipeWire exposes a pipewire pathname socket via SecurityContext.
+	EPipeWire
 	// EPulse copies the PulseAudio cookie to [hst.PrivateTmp] and exposes the PulseAudio socket.
 	EPulse
 
@@ -35,6 +37,8 @@ func (e Enablement) String() string {
 		return "x11"
 	case EDBus:
 		return "dbus"
+	case EPipeWire:
+		return "pipewire"
 	case EPulse:
 		return "pulseaudio"
 	default:
@@ -62,10 +66,11 @@ type Enablements Enablement
 
 // enablementsJSON is the [json] representation of [Enablements].
 type enablementsJSON = struct {
-	Wayland bool `json:"wayland,omitempty"`
-	X11     bool `json:"x11,omitempty"`
-	DBus    bool `json:"dbus,omitempty"`
-	Pulse   bool `json:"pulse,omitempty"`
+	Wayland  bool `json:"wayland,omitempty"`
+	X11      bool `json:"x11,omitempty"`
+	DBus     bool `json:"dbus,omitempty"`
+	PipeWire bool `json:"pipewire,omitempty"`
+	Pulse    bool `json:"pulse,omitempty"`
 }
 
 // Unwrap returns the underlying [Enablement].
@@ -81,10 +86,11 @@ func (e *Enablements) MarshalJSON() ([]byte, error) {
 		return nil, syscall.EINVAL
 	}
 	return json.Marshal(&enablementsJSON{
-		Wayland: Enablement(*e)&EWayland != 0,
-		X11:     Enablement(*e)&EX11 != 0,
-		DBus:    Enablement(*e)&EDBus != 0,
-		Pulse:   Enablement(*e)&EPulse != 0,
+		Wayland:  Enablement(*e)&EWayland != 0,
+		X11:      Enablement(*e)&EX11 != 0,
+		DBus:     Enablement(*e)&EDBus != 0,
+		PipeWire: Enablement(*e)&EPipeWire != 0,
+		Pulse:    Enablement(*e)&EPulse != 0,
 	})
 }
 
@@ -108,6 +114,9 @@ func (e *Enablements) UnmarshalJSON(data []byte) error {
 	if v.DBus {
 		ve |= EDBus
 	}
+	if v.PipeWire {
+		ve |= EPipeWire
+	}
 	if v.Pulse {
 		ve |= EPulse
 	}

hst/enablement_test.go

@@ -32,6 +32,7 @@ func TestEnablementString(t *testing.T) {
 		{hst.EWayland | hst.EDBus | hst.EPulse, "wayland, dbus, pulseaudio"},
 		{hst.EX11 | hst.EDBus | hst.EPulse, "x11, dbus, pulseaudio"},
 		{hst.EWayland | hst.EX11 | hst.EDBus | hst.EPulse, "wayland, x11, dbus, pulseaudio"},
+		{hst.EM - 1, "wayland, x11, dbus, pipewire, pulseaudio"},
 
 		{1 << 5, "e20"},
 		{1 << 6, "e40"},
@@ -62,8 +63,9 @@ func TestEnablements(t *testing.T) {
 		{"wayland", hst.NewEnablements(hst.EWayland), `{"wayland":true}`, `{"value":{"wayland":true},"magic":3236757504}`},
 		{"x11", hst.NewEnablements(hst.EX11), `{"x11":true}`, `{"value":{"x11":true},"magic":3236757504}`},
 		{"dbus", hst.NewEnablements(hst.EDBus), `{"dbus":true}`, `{"value":{"dbus":true},"magic":3236757504}`},
+		{"pipewire", hst.NewEnablements(hst.EPipeWire), `{"pipewire":true}`, `{"value":{"pipewire":true},"magic":3236757504}`},
 		{"pulse", hst.NewEnablements(hst.EPulse), `{"pulse":true}`, `{"value":{"pulse":true},"magic":3236757504}`},
-		{"all", hst.NewEnablements(hst.EWayland | hst.EX11 | hst.EDBus | hst.EPulse), `{"wayland":true,"x11":true,"dbus":true,"pulse":true}`, `{"value":{"wayland":true,"x11":true,"dbus":true,"pulse":true},"magic":3236757504}`},
+		{"all", hst.NewEnablements(hst.EM - 1), `{"wayland":true,"x11":true,"dbus":true,"pipewire":true,"pulse":true}`, `{"value":{"wayland":true,"x11":true,"dbus":true,"pipewire":true,"pulse":true},"magic":3236757504}`},
 	}
 
 	for _, tc := range testCases {

hst/fs.go

@@ -45,6 +45,9 @@ type Ops interface {
 	Root(host *check.Absolute, flags int) Ops
 	// Etc appends an op that expands host /etc into a toplevel symlink mirror with /etc semantics.
 	Etc(host *check.Absolute, prefix string) Ops
+
+	// Daemon appends an op that starts a daemon in the container and blocks until target appears.
+	Daemon(target, path *check.Absolute, args ...string) Ops
 }
 
 // ApplyState holds the address of [Ops] and any relevant application state.
@@ -124,6 +127,12 @@ func (f *FilesystemConfigJSON) MarshalJSON() ([]byte, error) {
 			*FSLink
 		}{fsType{FilesystemLink}, cv}
 
+	case *FSDaemon:
+		v = &struct {
+			fsType
+			*FSDaemon
+		}{fsType{FilesystemDaemon}, cv}
+
 	default:
 		return nil, FSImplError{f.FilesystemConfig}
 	}
@@ -152,6 +161,9 @@ func (f *FilesystemConfigJSON) UnmarshalJSON(data []byte) error {
 	case FilesystemLink:
 		*f = FilesystemConfigJSON{new(FSLink)}
 
+	case FilesystemDaemon:
+		*f = FilesystemConfigJSON{new(FSDaemon)}
+
 	default:
 		return FSTypeError(t.Type)
 	}

hst/fs_test.go

@@ -84,6 +84,16 @@ func TestFilesystemConfigJSON(t *testing.T) {
 		}, nil,
 			`{"type":"link","dst":"/run/current-system","linkname":"/run/current-system","dereference":true}`,
 			`{"fs":{"type":"link","dst":"/run/current-system","linkname":"/run/current-system","dereference":true},"magic":3236757504}`},
+
+		{"daemon", hst.FilesystemConfigJSON{
+			FilesystemConfig: &hst.FSDaemon{
+				Target: m("/run/user/1971/pulse/native"),
+				Exec:   m("/run/current-system/sw/bin/pipewire-pulse"),
+				Args:   []string{"-v"},
+			},
+		}, nil,
+			`{"type":"daemon","dst":"/run/user/1971/pulse/native","path":"/run/current-system/sw/bin/pipewire-pulse","args":["-v"]}`,
+			`{"fs":{"type":"daemon","dst":"/run/user/1971/pulse/native","path":"/run/current-system/sw/bin/pipewire-pulse","args":["-v"]},"magic":3236757504}`},
 	}
 
 	for _, tc := range testCases {
@@ -345,6 +355,10 @@ func (p opsAdapter) Etc(host *check.Absolute, prefix string) hst.Ops {
 	return opsAdapter{p.Ops.Etc(host, prefix)}
 }
 
+func (p opsAdapter) Daemon(target, path *check.Absolute, args ...string) hst.Ops {
+	return opsAdapter{p.Ops.Daemon(target, path, args...)}
+}
+
 func m(pathname string) *check.Absolute { return check.MustAbs(pathname) }
 func ms(pathnames ...string) []*check.Absolute {
 	as := make([]*check.Absolute, len(pathnames))

hst/fsdaemon.go

@@ -0,0 +1,48 @@
+package hst
+
+import (
+	"encoding/gob"
+
+	"hakurei.app/container/check"
+)
+
+func init() { gob.Register(new(FSDaemon)) }
+
+// FilesystemDaemon is the type string of a daemon.
+const FilesystemDaemon = "daemon"
+
+// FSDaemon represents a daemon to be started in the container.
+type FSDaemon struct {
+	// Pathname indicating readiness of daemon.
+	Target *check.Absolute `json:"dst"`
+	// Absolute pathname to daemon executable file.
+	Exec *check.Absolute `json:"path"`
+	// Arguments (excl. first) passed to daemon.
+	Args []string `json:"args"`
+}
+
+func (d *FSDaemon) Valid() bool { return d != nil && d.Target != nil && d.Exec != nil }
+
+func (d *FSDaemon) Path() *check.Absolute {
+	if !d.Valid() {
+		return nil
+	}
+	return d.Target
+}
+
+func (d *FSDaemon) Host() []*check.Absolute { return nil }
+
+func (d *FSDaemon) Apply(z *ApplyState) {
+	if !d.Valid() {
+		return
+	}
+	z.Daemon(d.Target, d.Exec, d.Args...)
+}
+
+func (d *FSDaemon) String() string {
+	if !d.Valid() {
+		return "<invalid>"
+	}
+
+	return "daemon:" + d.Target.String()
+}

hst/fsdaemon_test.go

@@ -0,0 +1,29 @@
+package hst_test
+
+import (
+	"testing"
+
+	"hakurei.app/container"
+	"hakurei.app/hst"
+)
+
+func TestFSDaemon(t *testing.T) {
+	t.Parallel()
+
+	checkFs(t, []fsTestCase{
+		{"nil", (*hst.FSDaemon)(nil), false, nil, nil, nil, "<invalid>"},
+		{"zero", new(hst.FSDaemon), false, nil, nil, nil, "<invalid>"},
+
+		{"pipewire-pulse", &hst.FSDaemon{
+			Target: m("/run/user/1971/pulse/native"),
+			Exec:   m("/run/current-system/sw/bin/pipewire-pulse"),
+			Args:   []string{"-v"},
+		}, true, container.Ops{
+			&container.DaemonOp{
+				Target: m("/run/user/1971/pulse/native"),
+				Path:   m("/run/current-system/sw/bin/pipewire-pulse"),
+				Args:   []string{"-v"},
+			},
+		}, m("/run/user/1971/pulse/native"), nil, `daemon:/run/user/1971/pulse/native`},
+	})
+}

hst/hst.go

@@ -54,6 +54,9 @@ type Paths struct {
 
 // Info holds basic system information collected from the implementation.
 type Info struct {
+	// WaylandVersion is the libwayland value of WAYLAND_VERSION.
+	WaylandVersion string `json:"WAYLAND_VERSION"`
+
 	// Version is a hardcoded version string.
 	Version string `json:"version"`
 	// User is the userid according to hsu.
@@ -67,7 +70,7 @@ func Template() *Config {
 	return &Config{
 		ID: "org.chromium.Chromium",
 
-		Enablements: NewEnablements(EWayland | EDBus | EPulse),
+		Enablements: NewEnablements(EWayland | EDBus | EPipeWire),
 
 		SessionBus: &BusConfig{
 			See: nil,
@@ -89,7 +92,6 @@ func Template() *Config {
 			Log:       false,
 			Filter:    true,
 		},
-		DirectWayland: false,
 
 		ExtraPerms: []ExtraPermConfig{
 			{Path: fhs.AbsVarLib.Append("hakurei/u0"), Ensure: true, Execute: true},

hst/hst_test.go

@@ -105,7 +105,7 @@ func TestTemplate(t *testing.T) {
 	"enablements": {
 		"wayland": true,
 		"dbus": true,
-		"pulse": true
+		"pipewire": true
 	},
 	"session_bus": {
 		"see": null,

hst/instance_test.go

@@ -6,11 +6,13 @@ import (
 	"reflect"
 	"testing"
 	"time"
-	_ "unsafe"
+	_ "unsafe" // for go:linkname
 
 	"hakurei.app/hst"
 )
 
+// Made available here to check time encoding behaviour of [hst.ID].
+//
 //go:linkname newInstanceID hakurei.app/hst.newInstanceID
 func newInstanceID(id *hst.ID, p uint64) error
 

internal/acl/acl_test.go

@@ -13,7 +13,7 @@ import (
 	"strconv"
 	"testing"
 
-	"hakurei.app/system/acl"
+	"hakurei.app/internal/acl"
 )
 
 const testFileName = "acl.test"

internal/acl/libacl-helper.c

@@ -0,0 +1,90 @@
+#include "libacl-helper.h"
+#include <acl/libacl.h>
+#include <stdbool.h>
+#include <stdlib.h>
+#include <sys/acl.h>
+
+int hakurei_acl_update_file_by_uid(const char *path_p, uid_t uid,
+                                   acl_perm_t *perms, size_t plen) {
+    int ret;
+    bool v;
+    int i;
+    acl_t acl;
+    acl_entry_t entry;
+    acl_tag_t tag_type;
+    void *qualifier_p;
+    acl_permset_t permset;
+
+    ret = -1; /* acl_get_file */
+    acl = acl_get_file(path_p, ACL_TYPE_ACCESS);
+    if (acl == NULL)
+        goto out;
+
+    /* prune entries by uid */
+    for (i = acl_get_entry(acl, ACL_FIRST_ENTRY, &entry); i == 1;
+         i = acl_get_entry(acl, ACL_NEXT_ENTRY, &entry)) {
+        ret = -2; /* acl_get_tag_type */
+        if (acl_get_tag_type(entry, &tag_type) != 0)
+            goto out;
+        if (tag_type != ACL_USER)
+            continue;
+
+        ret = -3; /* acl_get_qualifier */
+        qualifier_p = acl_get_qualifier(entry);
+        if (qualifier_p == NULL)
+            goto out;
+        v = *(uid_t *)qualifier_p == uid;
+        acl_free(qualifier_p);
+
+        if (!v)
+            continue;
+
+        ret = -4; /* acl_delete_entry */
+        if (acl_delete_entry(acl, entry) != 0)
+            goto out;
+    }
+
+    if (plen == 0)
+        goto set;
+
+    ret = -5; /* acl_create_entry */
+    if (acl_create_entry(&acl, &entry) != 0)
+        goto out;
+
+    ret = -6; /* acl_get_permset */
+    if (acl_get_permset(entry, &permset) != 0)
+        goto out;
+
+    ret = -7; /* acl_add_perm */
+    for (i = 0; i < plen; i++) {
+        if (acl_add_perm(permset, perms[i]) != 0)
+            goto out;
+    }
+
+    ret = -8; /* acl_set_tag_type */
+    if (acl_set_tag_type(entry, ACL_USER) != 0)
+        goto out;
+
+    ret = -9; /* acl_set_qualifier */
+    if (acl_set_qualifier(entry, (void *)&uid) != 0)
+        goto out;
+
+set:
+    ret = -10; /* acl_calc_mask */
+    if (acl_calc_mask(&acl) != 0)
+        goto out;
+
+    ret = -11; /* acl_valid */
+    if (acl_valid(acl) != 0)
+        goto out;
+
+    ret = -12; /* acl_set_file */
+    if (acl_set_file(path_p, ACL_TYPE_ACCESS, acl) == 0)
+        ret = 0;
+
+out:
+    free((void *)path_p);
+    if (acl != NULL)
+        acl_free((void *)acl);
+    return ret;
+}

internal/acl/perms_test.go

@@ -3,7 +3,7 @@ package acl_test
 import (
 	"testing"
 
-	"hakurei.app/system/acl"
+	"hakurei.app/internal/acl"
 )
 
 func TestPerms(t *testing.T) {

internal/dbus/address_test.go

@@ -5,7 +5,7 @@ import (
 	"reflect"
 	"testing"
 
-	"hakurei.app/system/dbus"
+	"hakurei.app/internal/dbus"
 )
 
 func TestParse(t *testing.T) {

internal/dbus/config_test.go

@@ -7,7 +7,7 @@ import (
 	"testing"
 
 	"hakurei.app/hst"
-	"hakurei.app/system/dbus"
+	"hakurei.app/internal/dbus"
 )
 
 func TestConfigArgs(t *testing.T) {

internal/dbus/dbus_test.go

@@ -11,9 +11,9 @@ import (
 	"testing"
 	"time"
 
-	"hakurei.app/helper"
+	"hakurei.app/internal/dbus"
+	"hakurei.app/internal/helper"
 	"hakurei.app/message"
-	"hakurei.app/system/dbus"
 )
 
 func TestFinalise(t *testing.T) {

internal/dbus/proc.go

@@ -12,7 +12,7 @@ import (
 	"hakurei.app/container/check"
 	"hakurei.app/container/seccomp"
 	"hakurei.app/container/std"
-	"hakurei.app/helper"
+	"hakurei.app/internal/helper"
 	"hakurei.app/ldd"
 )
 
@@ -54,7 +54,7 @@ func (p *Proxy) Start() error {
 		}
 
 		var libPaths []*check.Absolute
-		if entries, err := ldd.Exec(ctx, p.msg, toolPath.String()); err != nil {
+		if entries, err := ldd.Resolve(ctx, p.msg, toolPath); err != nil {
 			return err
 		} else {
 			libPaths = ldd.Path(entries)

internal/dbus/proc_test.go

@@ -5,7 +5,7 @@ import (
 	"testing"
 
 	"hakurei.app/container"
-	"hakurei.app/helper"
+	"hakurei.app/internal/helper"
 )
 
 func TestMain(m *testing.M) { container.TryArgv0(nil); helper.InternalHelperStub(); os.Exit(m.Run()) }

internal/dbus/proxy.go

@@ -6,8 +6,8 @@ import (
 	"sync"
 	"syscall"
 
-	"hakurei.app/helper"
 	"hakurei.app/hst"
+	"hakurei.app/internal/helper"
 	"hakurei.app/message"
 )
 

internal/helper/args_test.go

@@ -7,7 +7,7 @@ import (
 	"syscall"
 	"testing"
 
-	"hakurei.app/helper"
+	"hakurei.app/internal/helper"
 )
 
 func TestArgsString(t *testing.T) {

internal/helper/cmd.go

@@ -10,7 +10,7 @@ import (
 	"sync"
 	"syscall"
 
-	"hakurei.app/helper/proc"
+	"hakurei.app/internal/helper/proc"
 )
 
 // NewDirect initialises a new direct Helper instance with wt as the null-terminated argument writer.

internal/helper/cmd_test.go

@@ -9,7 +9,7 @@ import (
 	"testing"
 
 	"hakurei.app/container"
-	"hakurei.app/helper"
+	"hakurei.app/internal/helper"
 )
 
 func TestCmd(t *testing.T) {

internal/helper/container.go

@@ -11,7 +11,7 @@ import (
 
 	"hakurei.app/container"
 	"hakurei.app/container/check"
-	"hakurei.app/helper/proc"
+	"hakurei.app/internal/helper/proc"
 	"hakurei.app/message"
 )
 

internal/helper/container_test.go

@@ -9,7 +9,7 @@ import (
 	"hakurei.app/container"
 	"hakurei.app/container/check"
 	"hakurei.app/container/fhs"
-	"hakurei.app/helper"
+	"hakurei.app/internal/helper"
 )
 
 func TestContainer(t *testing.T) {

internal/helper/helper.go

@@ -8,7 +8,7 @@ import (
 	"os"
 	"time"
 
-	"hakurei.app/helper/proc"
+	"hakurei.app/internal/helper/proc"
 )
 
 var WaitDelay = 2 * time.Second

internal/helper/helper_test.go

@@ -13,7 +13,7 @@ import (
 	"testing"
 	"time"
 
-	"hakurei.app/helper"
+	"hakurei.app/internal/helper"
 )
 
 var (