release: 0.3.4

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

.gitea/workflows/test.yml

@@ -72,6 +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
@@ -96,6 +113,7 @@ jobs:
       - race
       - sandbox
       - sandbox-race
+      - sharefs
       - hpkg
     runs-on: nix
     steps:

.gitignore

@@ -27,6 +27,7 @@ go.work.sum
 
 # go generate
 /cmd/hakurei/LICENSE
+/internal/pkg/testdata/testtool
 
 # release
 /dist/hakurei-*

cmd/hakurei/command.go

@@ -91,7 +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 {
@@ -146,8 +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{
@@ -297,8 +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")
 	}
 
 	{

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/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"),
@@ -62,7 +62,7 @@ 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
@@ -159,7 +159,7 @@ 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
@@ -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/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/mbf/main.go

@@ -0,0 +1,243 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"log"
+	"os"
+	"os/signal"
+	"path/filepath"
+	"runtime"
+	"syscall"
+	"unique"
+
+	"hakurei.app/command"
+	"hakurei.app/container"
+	"hakurei.app/container/check"
+	"hakurei.app/internal/pkg"
+	"hakurei.app/internal/rosa"
+	"hakurei.app/message"
+)
+
+func main() {
+	container.TryArgv0(nil)
+
+	log.SetFlags(0)
+	log.SetPrefix("mbf: ")
+	msg := message.New(log.Default())
+
+	if os.Geteuid() == 0 {
+		log.Fatal("this program must not run as root")
+	}
+
+	var cache *pkg.Cache
+	ctx, stop := signal.NotifyContext(context.Background(),
+		syscall.SIGINT, syscall.SIGTERM, syscall.SIGHUP)
+	defer stop()
+	defer func() {
+		if cache != nil {
+			cache.Close()
+		}
+
+		if r := recover(); r != nil {
+			fmt.Println(r)
+			log.Fatal("consider scrubbing the on-disk cache")
+		}
+	}()
+
+	var (
+		flagQuiet  bool
+		flagCures  int
+		flagBase   string
+		flagTShift int
+	)
+	c := command.New(os.Stderr, log.Printf, "mbf", func([]string) (err error) {
+		msg.SwapVerbose(!flagQuiet)
+
+		var base *check.Absolute
+		if flagBase, err = filepath.Abs(flagBase); err != nil {
+			return
+		} else if base, err = check.NewAbs(flagBase); err != nil {
+			return
+		}
+		if cache, err = pkg.Open(ctx, msg, flagCures, base); err == nil {
+			if flagTShift < 0 {
+				cache.SetThreshold(0)
+			} else if flagTShift > 31 {
+				cache.SetThreshold(1 << 31)
+			} else {
+				cache.SetThreshold(1 << flagTShift)
+			}
+		}
+		return
+	}).Flag(
+		&flagQuiet,
+		"q", command.BoolFlag(false),
+		"Do not print cure messages",
+	).Flag(
+		&flagCures,
+		"cures", command.IntFlag(0),
+		"Maximum number of dependencies to cure at any given time",
+	).Flag(
+		&flagBase,
+		"d", command.StringFlag("cache"),
+		"Directory to store cured artifacts",
+	).Flag(
+		&flagTShift,
+		"tshift", command.IntFlag(-1),
+		"Dependency graph size exponent, to the power of 2",
+	)
+
+	{
+		var flagShifts int
+		c.NewCommand(
+			"scrub", "Examine the on-disk cache for errors",
+			func(args []string) error {
+				if len(args) > 0 {
+					return errors.New("scrub expects no arguments")
+				}
+				if flagShifts < 0 || flagShifts > 31 {
+					flagShifts = 12
+				}
+				return cache.Scrub(runtime.NumCPU() << flagShifts)
+			},
+		).Flag(
+			&flagShifts,
+			"shift", command.IntFlag(12),
+			"Scrub parallelism size exponent, to the power of 2",
+		)
+	}
+
+	c.NewCommand(
+		"stage3",
+		"Check for toolchain 3-stage non-determinism",
+		func(args []string) (err error) {
+			_, _, _, stage2 := (rosa.Std - 1).NewLLVM()
+			_, _, _, stage3 := rosa.Std.NewLLVM()
+			var (
+				pathname *check.Absolute
+				checksum [2]unique.Handle[pkg.Checksum]
+			)
+
+			if pathname, checksum[0], err = cache.Cure(stage2); err != nil {
+				return err
+			}
+			log.Println("stage2:", pathname)
+			if pathname, checksum[1], err = cache.Cure(stage3); err != nil {
+				return err
+			}
+			log.Println("stage3:", pathname)
+
+			if checksum[0] != checksum[1] {
+				err = &pkg.ChecksumMismatchError{
+					Got:  checksum[0].Value(),
+					Want: checksum[1].Value(),
+				}
+			}
+			return
+		},
+	)
+
+	c.NewCommand(
+		"cure",
+		"Cure the named artifact and show its path",
+		func(args []string) error {
+			if len(args) != 1 {
+				return errors.New("cure requires 1 argument")
+			}
+			var p rosa.PArtifact
+			switch args[0] {
+			case "acl":
+				p = rosa.ACL
+			case "attr":
+				p = rosa.Attr
+			case "autoconf":
+				p = rosa.Autoconf
+			case "bash":
+				p = rosa.Bash
+			case "busybox":
+				p = rosa.Busybox
+			case "cmake":
+				p = rosa.CMake
+			case "coreutils":
+				p = rosa.Coreutils
+			case "diffutils":
+				p = rosa.Diffutils
+			case "gettext":
+				p = rosa.Gettext
+			case "git":
+				p = rosa.Git
+			case "go":
+				p = rosa.Go
+			case "gperf":
+				p = rosa.Gperf
+			case "hakurei":
+				p = rosa.Hakurei
+			case "kernel-headers":
+				p = rosa.KernelHeaders
+			case "libXau":
+				p = rosa.LibXau
+			case "libexpat":
+				p = rosa.Libexpat
+			case "libseccomp":
+				p = rosa.Libseccomp
+			case "libxml2":
+				p = rosa.Libxml2
+			case "libffi":
+				p = rosa.Libffi
+			case "libgd":
+				p = rosa.Libgd
+			case "m4":
+				p = rosa.M4
+			case "make":
+				p = rosa.Make
+			case "meson":
+				p = rosa.Meson
+			case "ninja":
+				p = rosa.Ninja
+			case "patch":
+				p = rosa.Patch
+			case "perl":
+				p = rosa.Perl
+			case "pkg-config":
+				p = rosa.PkgConfig
+			case "python":
+				p = rosa.Python
+			case "rsync":
+				p = rosa.Rsync
+			case "setuptools":
+				p = rosa.Setuptools
+			case "wayland":
+				p = rosa.Wayland
+			case "wayland-protocols":
+				p = rosa.WaylandProtocols
+			case "xcb":
+				p = rosa.XCB
+			case "xcb-proto":
+				p = rosa.XCBProto
+			case "xproto":
+				p = rosa.Xproto
+			case "zlib":
+				p = rosa.Zlib
+
+			default:
+				return fmt.Errorf("unsupported artifact %q", args[0])
+			}
+
+			pathname, _, err := cache.Cure(rosa.Std.Load(p))
+			if err == nil {
+				log.Println(pathname)
+			}
+			return err
+
+		},
+	)
+
+	c.MustParse(os.Args[1:], func(err error) {
+		if cache != nil {
+			cache.Close()
+		}
+		log.Fatal(err)
+	})
+}

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/capability.go

@@ -14,6 +14,7 @@ const (
 
 	CAP_SYS_ADMIN    = 0x15
 	CAP_SETPCAP      = 0x8
+	CAP_NET_ADMIN    = 0xc
 	CAP_DAC_OVERRIDE = 0x1
 )
 

container/check/absolute.go

@@ -9,46 +9,60 @@ import (
 	"slices"
 	"strings"
 	"syscall"
+	"unique"
 )
 
 // AbsoluteError is returned by [NewAbs] and holds the invalid pathname.
-type AbsoluteError struct{ Pathname string }
+type AbsoluteError string
 
-func (e *AbsoluteError) Error() string { return fmt.Sprintf("path %q is not absolute", e.Pathname) }
-func (e *AbsoluteError) Is(target error) bool {
-	var ce *AbsoluteError
+func (e AbsoluteError) Error() string {
+	return fmt.Sprintf("path %q is not absolute", string(e))
+}
+
+func (e AbsoluteError) Is(target error) bool {
+	var ce AbsoluteError
 	if !errors.As(target, &ce) {
 		return errors.Is(target, syscall.EINVAL)
 	}
-	return *e == *ce
+	return e == ce
 }
 
 // Absolute holds a pathname checked to be absolute.
-type Absolute struct{ pathname string }
+type Absolute struct{ pathname unique.Handle[string] }
+
+// ok returns whether [Absolute] is not the zero value.
+func (a *Absolute) ok() bool { return a != nil && *a != (Absolute{}) }
 
 // unsafeAbs returns [check.Absolute] on any string value.
-func unsafeAbs(pathname string) *Absolute { return &Absolute{pathname} }
+func unsafeAbs(pathname string) *Absolute {
+	return &Absolute{unique.Make(pathname)}
+}
 
+// String returns the checked pathname.
 func (a *Absolute) String() string {
-	if a.pathname == "" {
+	if !a.ok() {
 		panic("attempted use of zero Absolute")
 	}
+	return a.pathname.Value()
+}
+
+// Handle returns the underlying [unique.Handle].
+func (a *Absolute) Handle() unique.Handle[string] {
 	return a.pathname
 }
 
+// Is efficiently compares the underlying pathname.
 func (a *Absolute) Is(v *Absolute) bool {
 	if a == nil && v == nil {
 		return true
 	}
-	return a != nil && v != nil &&
-		a.pathname != "" && v.pathname != "" &&
-		a.pathname == v.pathname
+	return a.ok() && v.ok() && a.pathname == v.pathname
 }
 
 // NewAbs checks pathname and returns a new [Absolute] if pathname is absolute.
 func NewAbs(pathname string) (*Absolute, error) {
 	if !path.IsAbs(pathname) {
-		return nil, &AbsoluteError{pathname}
+		return nil, AbsoluteError(pathname)
 	}
 	return unsafeAbs(pathname), nil
 }
@@ -70,35 +84,49 @@ func (a *Absolute) Append(elem ...string) *Absolute {
 // Dir calls [path.Dir] with [Absolute] as its argument.
 func (a *Absolute) Dir() *Absolute { return unsafeAbs(path.Dir(a.String())) }
 
-func (a *Absolute) GobEncode() ([]byte, error) { return []byte(a.String()), nil }
+// GobEncode returns the checked pathname.
+func (a *Absolute) GobEncode() ([]byte, error) {
+	return []byte(a.String()), nil
+}
+
+// GobDecode stores data if it represents an absolute pathname.
 func (a *Absolute) GobDecode(data []byte) error {
 	pathname := string(data)
 	if !path.IsAbs(pathname) {
-		return &AbsoluteError{pathname}
+		return AbsoluteError(pathname)
 	}
-	a.pathname = pathname
+	a.pathname = unique.Make(pathname)
 	return nil
 }
 
-func (a *Absolute) MarshalJSON() ([]byte, error) { return json.Marshal(a.String()) }
+// MarshalJSON returns a JSON representation of the checked pathname.
+func (a *Absolute) MarshalJSON() ([]byte, error) {
+	return json.Marshal(a.String())
+}
+
+// UnmarshalJSON stores data if it represents an absolute pathname.
 func (a *Absolute) UnmarshalJSON(data []byte) error {
 	var pathname string
 	if err := json.Unmarshal(data, &pathname); err != nil {
 		return err
 	}
 	if !path.IsAbs(pathname) {
-		return &AbsoluteError{pathname}
+		return AbsoluteError(pathname)
 	}
-	a.pathname = pathname
+	a.pathname = unique.Make(pathname)
 	return nil
 }
 
 // SortAbs calls [slices.SortFunc] for a slice of [Absolute].
 func SortAbs(x []*Absolute) {
-	slices.SortFunc(x, func(a, b *Absolute) int { return strings.Compare(a.String(), b.String()) })
+	slices.SortFunc(x, func(a, b *Absolute) int {
+		return strings.Compare(a.String(), b.String())
+	})
 }
 
 // CompactAbs calls [slices.CompactFunc] for a slice of [Absolute].
 func CompactAbs(s []*Absolute) []*Absolute {
-	return slices.CompactFunc(s, func(a *Absolute, b *Absolute) bool { return a.String() == b.String() })
+	return slices.CompactFunc(s, func(a *Absolute, b *Absolute) bool {
+		return a.Is(b)
+	})
 }

container/check/absolute_test.go

@@ -31,8 +31,8 @@ func TestAbsoluteError(t *testing.T) {
 	}{
 		{"EINVAL", new(AbsoluteError), syscall.EINVAL, true},
 		{"not EINVAL", new(AbsoluteError), syscall.EBADE, false},
-		{"ne val", new(AbsoluteError), &AbsoluteError{Pathname: "etc"}, false},
-		{"equals", &AbsoluteError{Pathname: "etc"}, &AbsoluteError{Pathname: "etc"}, true},
+		{"ne val", new(AbsoluteError), AbsoluteError("etc"), false},
+		{"equals", AbsoluteError("etc"), AbsoluteError("etc"), true},
 	}
 
 	for _, tc := range testCases {
@@ -45,7 +45,7 @@ func TestAbsoluteError(t *testing.T) {
 		t.Parallel()
 
 		want := `path "etc" is not absolute`
-		if got := (&AbsoluteError{Pathname: "etc"}).Error(); got != want {
+		if got := (AbsoluteError("etc")).Error(); got != want {
 			t.Errorf("Error: %q, want %q", got, want)
 		}
 	})
@@ -62,8 +62,8 @@ func TestNewAbs(t *testing.T) {
 		wantErr  error
 	}{
 		{"good", "/etc", MustAbs("/etc"), nil},
-		{"not absolute", "etc", nil, &AbsoluteError{Pathname: "etc"}},
-		{"zero", "", nil, &AbsoluteError{Pathname: ""}},
+		{"not absolute", "etc", nil, AbsoluteError("etc")},
+		{"zero", "", nil, AbsoluteError("")},
 	}
 
 	for _, tc := range testCases {
@@ -84,7 +84,7 @@ func TestNewAbs(t *testing.T) {
 		t.Parallel()
 
 		defer func() {
-			wantPanic := &AbsoluteError{Pathname: "etc"}
+			wantPanic := AbsoluteError("etc")
 
 			if r := recover(); !reflect.DeepEqual(r, wantPanic) {
 				t.Errorf("MustAbs: panic = %v; want %v", r, wantPanic)
@@ -175,7 +175,7 @@ func TestCodecAbsolute(t *testing.T) {
 
 			`"/etc"`, `{"val":"/etc","magic":3236757504}`},
 		{"not absolute", nil,
-			&AbsoluteError{Pathname: "etc"},
+			AbsoluteError("etc"),
 			"\t\x7f\x05\x01\x02\xff\x82\x00\x00\x00\a\xff\x80\x00\x03etc",
 			",\xff\x83\x03\x01\x01\x06sCheck\x01\xff\x84\x00\x01\x02\x01\bPathname\x01\xff\x80\x00\x01\x05Magic\x01\x06\x00\x00\x00\t\x7f\x05\x01\x02\xff\x82\x00\x00\x00\x0f\xff\x84\x01\x03etc\x01\xfb\x01\x81\xda\x00\x00\x00",
 

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,
 
@@ -262,12 +263,17 @@ func (p *Container) Start() error {
 			CAP_SYS_ADMIN,
 			// drop capabilities
 			CAP_SETPCAP,
+			// bring up loopback interface
+			CAP_NET_ADMIN,
 			// overlay access to upperdir and workdir
 			CAP_DAC_OVERRIDE,
 		},
 
 		UseCgroupFD: p.Cgroup != nil,
 	}
+	if !p.AllowOrphan {
+		p.cmd.SysProcAttr.Pdeathsig = SIGKILL
+	}
 	if p.cmd.SysProcAttr.UseCgroupFD {
 		p.cmd.SysProcAttr.CgroupFD = *p.Cgroup
 	}

container/dispatcher.go

@@ -61,6 +61,8 @@ type syscallDispatcher interface {
 	mountTmpfs(fsname, target string, flags uintptr, size int, perm os.FileMode) error
 	// ensureFile provides ensureFile.
 	ensureFile(name string, perm, pperm os.FileMode) error
+	// mustLoopback provides mustLoopback.
+	mustLoopback(msg message.Msg)
 
 	// seccompLoad provides [seccomp.Load].
 	seccompLoad(rules []std.NativeRule, flags seccomp.ExportFlag) error
@@ -164,6 +166,7 @@ func (k direct) mountTmpfs(fsname, target string, flags uintptr, size int, perm
 func (direct) ensureFile(name string, perm, pperm os.FileMode) error {
 	return ensureFile(name, perm, pperm)
 }
+func (direct) mustLoopback(msg message.Msg) { mustLoopback(msg) }
 
 func (direct) seccompLoad(rules []std.NativeRule, flags seccomp.ExportFlag) error {
 	return seccomp.Load(rules, flags)

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()
@@ -457,6 +465,8 @@ func (k *kstub) ensureFile(name string, perm, pperm os.FileMode) error {
 		stub.CheckArg(k.Stub, "pperm", pperm, 2))
 }
 
+func (*kstub) mustLoopback(message.Msg) { /* noop */ }
+
 func (k *kstub) seccompLoad(rules []std.NativeRule, flags seccomp.ExportFlag) error {
 	k.Helper()
 	return k.Expects("seccompLoad").Error(
@@ -472,6 +482,10 @@ 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 +773,8 @@ 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 = messagePrefix[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/errors_test.go

@@ -37,7 +37,7 @@ func TestMessageFromError(t *testing.T) {
 			Err:  stub.UniqueError(0xdeadbeef),
 		}, "cannot mount /sysroot: unique error 3735928559 injected by the test suite", true},
 
-		{"absolute", &check.AbsoluteError{Pathname: "etc/mtab"},
+		{"absolute", check.AbsoluteError("etc/mtab"),
 			`path "etc/mtab" is not absolute`, true},
 
 		{"repeat", OpRepeatError("autoetc"),

container/fhs/abs.go

@@ -26,6 +26,8 @@ var (
 	// AbsRunUser is [RunUser] as [check.Absolute].
 	AbsRunUser = unsafeAbs(RunUser)
 
+	// AbsUsr is [Usr] as [check.Absolute].
+	AbsUsr = unsafeAbs(Usr)
 	// AbsUsrBin is [UsrBin] as [check.Absolute].
 	AbsUsrBin = unsafeAbs(UsrBin)
 

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) }
 
@@ -143,6 +170,10 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 		offsetSetup = int(setupFd + 1)
 	}
 
+	if !params.HostNet {
+		k.mustLoopback(msg)
+	}
+
 	// write uid/gid map here so parent does not need to set dumpable
 	if err := k.setDumpable(SUID_DUMP_USER); err != nil {
 		k.fatalf(msg, "cannot set SUID_DUMP_USER: %v", err)
@@ -180,7 +211,9 @@ 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 +363,97 @@ 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 +465,11 @@ 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 +479,7 @@ 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 +511,9 @@ 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/initoverlay_test.go

@@ -312,7 +312,10 @@ func TestMountOverlayOp(t *testing.T) {
 			},
 		}},
 
-		{"ephemeral", new(Ops).OverlayEphemeral(check.MustAbs("/nix/store"), check.MustAbs("/mnt-root/nix/.ro-store")), Ops{
+		{"ephemeral", new(Ops).OverlayEphemeral(
+			check.MustAbs("/nix/store"),
+			check.MustAbs("/mnt-root/nix/.ro-store"),
+		), Ops{
 			&MountOverlayOp{
 				Target: check.MustAbs("/nix/store"),
 				Lower:  []*check.Absolute{check.MustAbs("/mnt-root/nix/.ro-store")},
@@ -320,7 +323,10 @@ func TestMountOverlayOp(t *testing.T) {
 			},
 		}},
 
-		{"readonly", new(Ops).OverlayReadonly(check.MustAbs("/nix/store"), check.MustAbs("/mnt-root/nix/.ro-store")), Ops{
+		{"readonly", new(Ops).OverlayReadonly(
+			check.MustAbs("/nix/store"),
+			check.MustAbs("/mnt-root/nix/.ro-store"),
+		), Ops{
 			&MountOverlayOp{
 				Target: check.MustAbs("/nix/store"),
 				Lower:  []*check.Absolute{check.MustAbs("/mnt-root/nix/.ro-store")},

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

@@ -31,7 +31,7 @@ func (l *SymlinkOp) Valid() bool { return l != nil && l.Target != nil && l.LinkN
 func (l *SymlinkOp) early(_ *setupState, k syscallDispatcher) error {
 	if l.Dereference {
 		if !path.IsAbs(l.LinkName) {
-			return &check.AbsoluteError{Pathname: l.LinkName}
+			return check.AbsoluteError(l.LinkName)
 		}
 		if name, err := k.readlink(l.LinkName); err != nil {
 			return err
@@ -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/initsymlink_test.go

@@ -23,7 +23,7 @@ func TestSymlinkOp(t *testing.T) {
 			Target:      check.MustAbs("/etc/mtab"),
 			LinkName:    "etc/mtab",
 			Dereference: true,
-		}, nil, &check.AbsoluteError{Pathname: "etc/mtab"}, nil, nil},
+		}, nil, check.AbsoluteError("etc/mtab"), nil, nil},
 
 		{"readlink", &Params{ParentPerm: 0755}, &SymlinkOp{
 			Target:      check.MustAbs("/etc/mtab"),

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/netlink.go

@@ -0,0 +1,269 @@
+package container
+
+import (
+	"encoding/binary"
+	"errors"
+	"net"
+	"os"
+	. "syscall"
+	"unsafe"
+
+	"hakurei.app/container/std"
+	"hakurei.app/message"
+)
+
+// rtnetlink represents a NETLINK_ROUTE socket.
+type rtnetlink struct {
+	// Sent as part of rtnetlink messages.
+	pid uint32
+	// AF_NETLINK socket.
+	fd int
+	// Whether the socket is open.
+	ok bool
+	// Message sequence number.
+	seq uint32
+}
+
+// open creates the underlying NETLINK_ROUTE socket.
+func (s *rtnetlink) open() (err error) {
+	if s.ok || s.fd < 0 {
+		return os.ErrInvalid
+	}
+
+	s.pid = uint32(Getpid())
+	if s.fd, err = Socket(
+		AF_NETLINK,
+		SOCK_RAW|SOCK_CLOEXEC,
+		NETLINK_ROUTE,
+	); err != nil {
+		return os.NewSyscallError("socket", err)
+	} else if err = Bind(s.fd, &SockaddrNetlink{
+		Family: AF_NETLINK,
+		Pid:    s.pid,
+	}); err != nil {
+		_ = s.close()
+		return os.NewSyscallError("bind", err)
+	} else {
+		s.ok = true
+		return nil
+	}
+}
+
+// close closes the underlying NETLINK_ROUTE socket.
+func (s *rtnetlink) close() error {
+	if !s.ok {
+		return os.ErrInvalid
+	}
+
+	s.ok = false
+	err := Close(s.fd)
+	s.fd = -1
+	return err
+}
+
+// roundtrip sends a netlink message and handles the reply.
+func (s *rtnetlink) roundtrip(data []byte) error {
+	if !s.ok {
+		return os.ErrInvalid
+	}
+
+	defer func() { s.seq++ }()
+
+	if err := Sendto(s.fd, data, 0, &SockaddrNetlink{
+		Family: AF_NETLINK,
+	}); err != nil {
+		return os.NewSyscallError("sendto", err)
+	}
+	buf := make([]byte, Getpagesize())
+
+done:
+	for {
+		p := buf
+		if n, _, err := Recvfrom(s.fd, p, 0); err != nil {
+			return os.NewSyscallError("recvfrom", err)
+		} else if n < NLMSG_HDRLEN {
+			return errors.ErrUnsupported
+		} else {
+			p = p[:n]
+		}
+
+		if msgs, err := ParseNetlinkMessage(p); err != nil {
+			return err
+		} else {
+			for _, m := range msgs {
+				if m.Header.Seq != s.seq || m.Header.Pid != s.pid {
+					return errors.ErrUnsupported
+				}
+				if m.Header.Type == NLMSG_DONE {
+					break done
+				}
+				if m.Header.Type == NLMSG_ERROR {
+					if len(m.Data) >= 4 {
+						errno := Errno(-std.ScmpInt(binary.NativeEndian.Uint32(m.Data)))
+						if errno == 0 {
+							return nil
+						}
+						return errno
+					}
+					return errors.ErrUnsupported
+				}
+			}
+		}
+	}
+
+	return nil
+}
+
+// mustRoundtrip calls roundtrip and terminates via msg for a non-nil error.
+func (s *rtnetlink) mustRoundtrip(msg message.Msg, data []byte) {
+	err := s.roundtrip(data)
+	if err == nil {
+		return
+	}
+	if closeErr := Close(s.fd); closeErr != nil {
+		msg.Verbosef("cannot close: %v", err)
+	}
+
+	switch err.(type) {
+	case *os.SyscallError:
+		msg.GetLogger().Fatalf("cannot %v", err)
+
+	case Errno:
+		msg.GetLogger().Fatalf("RTNETLINK answers: %v", err)
+
+	default:
+		msg.GetLogger().Fatalln("RTNETLINK answers with unexpected message")
+	}
+}
+
+// newaddrLo represents a RTM_NEWADDR message with two addresses.
+type newaddrLo struct {
+	header NlMsghdr
+	data   IfAddrmsg
+
+	r0 RtAttr
+	a0 [4]byte // in_addr
+	r1 RtAttr
+	a1 [4]byte // in_addr
+}
+
+// sizeofNewaddrLo is the expected size of newaddrLo.
+const sizeofNewaddrLo = NLMSG_HDRLEN + SizeofIfAddrmsg + (SizeofRtAttr+4)*2
+
+// newaddrLo returns the address of a populated newaddrLo.
+func (s *rtnetlink) newaddrLo(lo int) *newaddrLo {
+	return &newaddrLo{NlMsghdr{
+		Len:   sizeofNewaddrLo,
+		Type:  RTM_NEWADDR,
+		Flags: NLM_F_REQUEST | NLM_F_ACK | NLM_F_CREATE | NLM_F_EXCL,
+		Seq:   s.seq,
+		Pid:   s.pid,
+	}, IfAddrmsg{
+		Family:    AF_INET,
+		Prefixlen: 8,
+		Flags:     IFA_F_PERMANENT,
+		Scope:     RT_SCOPE_HOST,
+		Index:     uint32(lo),
+	}, RtAttr{
+		Len:  uint16(SizeofRtAttr + len(newaddrLo{}.a0)),
+		Type: IFA_LOCAL,
+	}, [4]byte{127, 0, 0, 1}, RtAttr{
+		Len:  uint16(SizeofRtAttr + len(newaddrLo{}.a1)),
+		Type: IFA_ADDRESS,
+	}, [4]byte{127, 0, 0, 1}}
+}
+
+func (msg *newaddrLo) toWireFormat() []byte {
+	var buf [sizeofNewaddrLo]byte
+
+	*(*uint32)(unsafe.Pointer(&buf[0:4][0])) = msg.header.Len
+	*(*uint16)(unsafe.Pointer(&buf[4:6][0])) = msg.header.Type
+	*(*uint16)(unsafe.Pointer(&buf[6:8][0])) = msg.header.Flags
+	*(*uint32)(unsafe.Pointer(&buf[8:12][0])) = msg.header.Seq
+	*(*uint32)(unsafe.Pointer(&buf[12:16][0])) = msg.header.Pid
+
+	buf[16] = msg.data.Family
+	buf[17] = msg.data.Prefixlen
+	buf[18] = msg.data.Flags
+	buf[19] = msg.data.Scope
+	*(*uint32)(unsafe.Pointer(&buf[20:24][0])) = msg.data.Index
+
+	*(*uint16)(unsafe.Pointer(&buf[24:26][0])) = msg.r0.Len
+	*(*uint16)(unsafe.Pointer(&buf[26:28][0])) = msg.r0.Type
+	copy(buf[28:32], msg.a0[:])
+	*(*uint16)(unsafe.Pointer(&buf[32:34][0])) = msg.r1.Len
+	*(*uint16)(unsafe.Pointer(&buf[34:36][0])) = msg.r1.Type
+	copy(buf[36:40], msg.a1[:])
+
+	return buf[:]
+}
+
+// newlinkLo represents a RTM_NEWLINK message.
+type newlinkLo struct {
+	header NlMsghdr
+	data   IfInfomsg
+}
+
+// sizeofNewlinkLo is the expected size of newlinkLo.
+const sizeofNewlinkLo = NLMSG_HDRLEN + SizeofIfInfomsg
+
+// newlinkLo returns the address of a populated newlinkLo.
+func (s *rtnetlink) newlinkLo(lo int) *newlinkLo {
+	return &newlinkLo{NlMsghdr{
+		Len:   sizeofNewlinkLo,
+		Type:  RTM_NEWLINK,
+		Flags: NLM_F_REQUEST | NLM_F_ACK,
+		Seq:   s.seq,
+		Pid:   s.pid,
+	}, IfInfomsg{
+		Family: AF_UNSPEC,
+		Index:  int32(lo),
+		Flags:  IFF_UP,
+		Change: IFF_UP,
+	}}
+}
+
+func (msg *newlinkLo) toWireFormat() []byte {
+	var buf [sizeofNewlinkLo]byte
+
+	*(*uint32)(unsafe.Pointer(&buf[0:4][0])) = msg.header.Len
+	*(*uint16)(unsafe.Pointer(&buf[4:6][0])) = msg.header.Type
+	*(*uint16)(unsafe.Pointer(&buf[6:8][0])) = msg.header.Flags
+	*(*uint32)(unsafe.Pointer(&buf[8:12][0])) = msg.header.Seq
+	*(*uint32)(unsafe.Pointer(&buf[12:16][0])) = msg.header.Pid
+
+	buf[16] = msg.data.Family
+	*(*uint16)(unsafe.Pointer(&buf[18:20][0])) = msg.data.Type
+	*(*int32)(unsafe.Pointer(&buf[20:24][0])) = msg.data.Index
+	*(*uint32)(unsafe.Pointer(&buf[24:28][0])) = msg.data.Flags
+	*(*uint32)(unsafe.Pointer(&buf[28:32][0])) = msg.data.Change
+
+	return buf[:]
+}
+
+// mustLoopback creates the loopback address and brings the lo interface up.
+// mustLoopback calls a fatal method of the underlying [log.Logger] of m with a
+// user-facing error message if RTNETLINK behaves unexpectedly.
+func mustLoopback(msg message.Msg) {
+	log := msg.GetLogger()
+
+	var lo int
+	if ifi, err := net.InterfaceByName("lo"); err != nil {
+		log.Fatalln(err)
+	} else {
+		lo = ifi.Index
+	}
+
+	var s rtnetlink
+	if err := s.open(); err != nil {
+		log.Fatalln(err)
+	}
+	defer func() {
+		if err := s.close(); err != nil {
+			msg.Verbosef("cannot close netlink: %v", err)
+		}
+	}()
+
+	s.mustRoundtrip(msg, s.newaddrLo(lo).toWireFormat())
+	s.mustRoundtrip(msg, s.newlinkLo(lo).toWireFormat())
+}

container/netlink_test.go

@@ -0,0 +1,72 @@
+package container
+
+import (
+	"testing"
+	"unsafe"
+)
+
+func TestSizeof(t *testing.T) {
+	if got := unsafe.Sizeof(newaddrLo{}); got != sizeofNewaddrLo {
+		t.Fatalf("newaddrLo: sizeof = %#x, want %#x", got, sizeofNewaddrLo)
+	}
+
+	if got := unsafe.Sizeof(newlinkLo{}); got != sizeofNewlinkLo {
+		t.Fatalf("newlinkLo: sizeof = %#x, want %#x", got, sizeofNewlinkLo)
+	}
+}
+
+func TestRtnetlinkMessage(t *testing.T) {
+	t.Parallel()
+
+	testCases := []struct {
+		name string
+		msg  interface{ toWireFormat() []byte }
+		want []byte
+	}{
+		{"newaddrLo", (&rtnetlink{pid: 1, seq: 0}).newaddrLo(1), []byte{
+			/* Len   */ 0x28, 0, 0, 0,
+			/* Type  */ 0x14, 0,
+			/* Flags */ 5, 6,
+			/* Seq   */ 0, 0, 0, 0,
+			/* Pid   */ 1, 0, 0, 0,
+
+			/* Family    */ 2,
+			/* Prefixlen */ 8,
+			/* Flags     */ 0x80,
+			/* Scope     */ 0xfe,
+			/* Index     */ 1, 0, 0, 0,
+
+			/* Len     */ 8, 0,
+			/* Type    */ 2, 0,
+			/* in_addr */ 127, 0, 0, 1,
+
+			/* Len     */ 8, 0,
+			/* Type    */ 1, 0,
+			/* in_addr */ 127, 0, 0, 1,
+		}},
+
+		{"newlinkLo", (&rtnetlink{pid: 1, seq: 1}).newlinkLo(1), []byte{
+			/* Len   */ 0x20, 0, 0, 0,
+			/* Type  */ 0x10, 0,
+			/* Flags */ 5, 0,
+			/* Seq   */ 1, 0, 0, 0,
+			/* Pid   */ 1, 0, 0, 0,
+
+			/* Family */ 0,
+			/* pad    */ 0,
+			/* Type   */ 0, 0,
+			/* Index  */ 1, 0, 0, 0,
+			/* Flags  */ 1, 0, 0, 0,
+			/* Change */ 1, 0, 0, 0,
+		}},
+	}
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Parallel()
+
+			if got := tc.msg.toWireFormat(); string(got) != string(tc.want) {
+				t.Fatalf("toWireFormat: %#v, want %#v", got, tc.want)
+			}
+		})
+	}
+}

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

@@ -1,12 +1,12 @@
 #!/bin/sh
 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/hakurei" "${DESTDIR}/usr/bin/hakurei"
+install -vDm0755 "bin/sharefs" "${DESTDIR}/usr/bin/sharefs"
 
-install -vDm4511 "bin/hsu" "${HAKUREI_INSTALL_PREFIX}/usr/bin/hsu"
-if [ ! -f "${HAKUREI_INSTALL_PREFIX}/etc/hsurc" ]; then
-    install -vDm0400 "hsurc.default" "${HAKUREI_INSTALL_PREFIX}/etc/hsurc"
+install -vDm4511 "bin/hsu" "${DESTDIR}/usr/bin/hsu"
+if [ ! -f "${DESTDIR}/etc/hsurc" ]; then
+    install -vDm0400 "hsurc.default" "${DESTDIR}/etc/hsurc"
 fi
 
-install -vDm0644 "comp/_hakurei" "${HAKUREI_INSTALL_PREFIX}/usr/share/zsh/site-functions/_hakurei"
+install -vDm0644 "comp/_hakurei" "${DESTDIR}/usr/share/zsh/site-functions/_hakurei"

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 ]; } ''
@@ -136,6 +138,10 @@
               ;
           };
           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 "${hakurei.name}-dist" { buildInputs = hakurei.targetPkgs ++ [ pkgs.pkgsStatic.musl ]; } ''
             # go requires XDG_CACHE_HOME for the build cache
@@ -160,7 +166,10 @@
           pkgs = nixpkgsFor.${system};
         in
         {
-          default = pkgs.mkShell { buildInputs = hakurei.targetPkgs; };
+          default = pkgs.mkShell {
+            buildInputs = hakurei.targetPkgs;
+            hardeningDisable = [ "fortify" ];
+          };
           withPackage = pkgs.mkShell { buildInputs = [ hakurei ] ++ hakurei.targetPkgs; };
 
           vm =
@@ -185,13 +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; };
                         };
                     };
                   }

go.mod

@@ -1,3 +1,3 @@
 module hakurei.app
 
-go 1.24
+go 1.25

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

@@ -70,7 +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,
@@ -92,7 +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,

internal/acl/acl_test.go

@@ -25,8 +25,7 @@ var (
 
 func TestUpdate(t *testing.T) {
 	if os.Getenv("GO_TEST_SKIP_ACL") == "1" {
-		t.Log("acl test skipped")
-		t.SkipNow()
+		t.Skip("acl test skipped")
 	}
 
 	testFilePath := path.Join(t.TempDir(), testFileName)
@@ -143,6 +142,7 @@ func (c *getFAclInvocation) run(name string) error {
 	}
 
 	c.cmd = exec.Command("getfacl", "--omit-header", "--absolute-names", "--numeric", name)
+	c.cmd.Stderr = os.Stderr
 
 	scanErr := make(chan error, 1)
 	if p, err := c.cmd.StdoutPipe(); err != nil {
@@ -254,7 +254,7 @@ func getfacl(t *testing.T, name string) []*getFAclResp {
 		t.Fatalf("getfacl: error = %v", err)
 	}
 	if len(c.pe) != 0 {
-		t.Errorf("errors encountered parsing getfacl output\n%s", errors.Join(c.pe...).Error())
+		t.Errorf("errors encountered parsing getfacl output\n%s", errors.Join(c.pe...))
 	}
 	return c.val
 }

internal/outcome/dispatcher.go

@@ -53,6 +53,10 @@ type syscallDispatcher interface {
 	readdir(name string) ([]os.DirEntry, error)
 	// tempdir provides [os.TempDir].
 	tempdir() string
+	// mkdir provides [os.Mkdir].
+	mkdir(name string, perm os.FileMode) error
+	// removeAll provides [os.RemoveAll].
+	removeAll(path string) error
 	// exit provides [os.Exit].
 	exit(code int)
 
@@ -62,6 +66,8 @@ type syscallDispatcher interface {
 	// lookupGroupId calls [user.LookupGroup] and returns the Gid field of the resulting [user.Group] struct.
 	lookupGroupId(name string) (string, error)
 
+	// lookPath provides exec.LookPath.
+	lookPath(file string) (string, error)
 	// cmdOutput provides the Output method of [exec.Cmd].
 	cmdOutput(cmd *exec.Cmd) ([]byte, error)
 
@@ -121,6 +127,8 @@ func (direct) stat(name string) (os.FileInfo, error)      { return os.Stat(name)
 func (direct) open(name string) (osFile, error)           { return os.Open(name) }
 func (direct) readdir(name string) ([]os.DirEntry, error) { return os.ReadDir(name) }
 func (direct) tempdir() string                            { return os.TempDir() }
+func (direct) mkdir(name string, perm os.FileMode) error  { return os.Mkdir(name, perm) }
+func (direct) removeAll(path string) error                { return os.RemoveAll(path) }
 func (direct) exit(code int)                              { os.Exit(code) }
 
 func (direct) evalSymlinks(path string) (string, error) { return filepath.EvalSymlinks(path) }
@@ -134,6 +142,7 @@ func (direct) lookupGroupId(name string) (gid string, err error) {
 	return
 }
 
+func (direct) lookPath(file string) (string, error)    { return exec.LookPath(file) }
 func (direct) cmdOutput(cmd *exec.Cmd) ([]byte, error) { return cmd.Output() }
 
 func (direct) notifyContext(parent context.Context, signals ...os.Signal) (ctx context.Context, stop context.CancelFunc) {

internal/outcome/dispatcher_test.go

@@ -701,10 +701,13 @@ func (panicDispatcher) stat(string) (os.FileInfo, error)                    { pa
 func (panicDispatcher) open(string) (osFile, error)                         { panic("unreachable") }
 func (panicDispatcher) readdir(string) ([]os.DirEntry, error)               { panic("unreachable") }
 func (panicDispatcher) tempdir() string                                     { panic("unreachable") }
+func (panicDispatcher) mkdir(string, os.FileMode) error                     { panic("unreachable") }
+func (panicDispatcher) removeAll(string) error                              { panic("unreachable") }
 func (panicDispatcher) exit(int)                                            { panic("unreachable") }
 func (panicDispatcher) evalSymlinks(string) (string, error)                 { panic("unreachable") }
 func (panicDispatcher) prctl(uintptr, uintptr, uintptr) error               { panic("unreachable") }
 func (panicDispatcher) lookupGroupId(string) (string, error)                { panic("unreachable") }
+func (panicDispatcher) lookPath(string) (string, error)                     { panic("unreachable") }
 func (panicDispatcher) cmdOutput(*exec.Cmd) ([]byte, error)                 { panic("unreachable") }
 func (panicDispatcher) overflowUid(message.Msg) int                         { panic("unreachable") }
 func (panicDispatcher) overflowGid(message.Msg) int                         { panic("unreachable") }

internal/outcome/outcome.go

@@ -70,7 +70,7 @@ type outcomeState struct {
 	// Copied from their respective exported values.
 	mapuid, mapgid *stringPair[int]
 
-	// Copied from [EnvPaths] per-process.
+	// Copied from [env.Paths] per-process.
 	sc hst.Paths
 	*env.Paths
 
@@ -172,6 +172,10 @@ type outcomeStateSys struct {
 
 	// Copied from [hst.Config]. Safe for read by spWaylandOp.toSystem only.
 	directWayland bool
+	// Copied from [hst.Config]. Safe for read by spPipeWireOp.toSystem only.
+	directPipeWire bool
+	// Copied from [hst.Config]. Safe for read by spPulseOp.toSystem only.
+	directPulse bool
 	// Copied header from [hst.Config]. Safe for read by spFilesystemOp.toSystem only.
 	extraPerms []hst.ExtraPermConfig
 	// Copied address from [hst.Config]. Safe for read by spDBusOp.toSystem only.
@@ -185,8 +189,8 @@ type outcomeStateSys struct {
 func (s *outcomeState) newSys(config *hst.Config, sys *system.I) *outcomeStateSys {
 	return &outcomeStateSys{
 		appId: config.ID, et: config.Enablements.Unwrap(),
-		directWayland: config.DirectWayland, extraPerms: config.ExtraPerms,
-		sessionBus: config.SessionBus, systemBus: config.SystemBus,
+		directWayland: config.DirectWayland, directPipeWire: config.DirectPipeWire, directPulse: config.DirectPulse,
+		extraPerms: config.ExtraPerms, sessionBus: config.SessionBus, systemBus: config.SystemBus,
 		sys: sys, outcomeState: s,
 	}
 }
@@ -253,6 +257,10 @@ type outcomeStateParams struct {
 	// Populated by spRuntimeOp.
 	runtimeDir *check.Absolute
 
+	// Path to pipewire-pulse server.
+	// Populated by spPipeWireOp if DirectPipeWire is false.
+	pipewirePulsePath *check.Absolute
+
 	as hst.ApplyState
 	*outcomeState
 }
@@ -292,6 +300,7 @@ func (state *outcomeStateSys) toSystem() error {
 		// optional via enablements
 		&spWaylandOp{},
 		&spX11Op{},
+		&spPipeWireOp{},
 		&spPulseOp{},
 		&spDBusOp{},
 

internal/outcome/run_test.go

@@ -27,7 +27,7 @@ import (
 	"hakurei.app/message"
 )
 
-func TestOutcomeMain(t *testing.T) {
+func TestOutcomeRun(t *testing.T) {
 	t.Parallel()
 	msg := message.New(nil)
 	msg.SwapVerbose(testing.Verbose())
@@ -67,18 +67,12 @@ func TestOutcomeMain(t *testing.T) {
 				"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa",
 			).
 
-			// ensureRuntimeDir
-			Ensure(m("/run/user/1971"), 0700).
-			UpdatePermType(system.User, m("/run/user/1971"), acl.Execute).
-			Ensure(m("/run/user/1971/hakurei"), 0700).
-			UpdatePermType(system.User, m("/run/user/1971/hakurei"), acl.Execute).
-
-			// runtime
-			Ephemeral(system.Process, m("/run/user/1971/hakurei/aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"), 0700).
-			UpdatePerm(m("/run/user/1971/hakurei/aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"), acl.Execute).
-
-			// spPulseOp
-			Link(m("/run/user/1971/pulse/native"), m("/run/user/1971/hakurei/aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa/pulse")).
+			// spPipeWireOp
+			PipeWire(
+				m("/tmp/hakurei.0/aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa/pipewire"),
+				"org.chromium.Chromium",
+				"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa",
+			).
 
 			// spDBusOp
 			MustProxyDBus(
@@ -106,8 +100,6 @@ func TestOutcomeMain(t *testing.T) {
 				"GOOGLE_DEFAULT_CLIENT_ID=77185425430.apps.googleusercontent.com",
 				"GOOGLE_DEFAULT_CLIENT_SECRET=OTJgUOQcT7lO7GsGZq2G4IlT",
 				"HOME=/data/data/org.chromium.Chromium",
-				"PULSE_COOKIE=/.hakurei/pulse-cookie",
-				"PULSE_SERVER=unix:/run/user/1971/pulse/native",
 				"SHELL=/run/current-system/sw/bin/zsh",
 				"TERM=xterm-256color",
 				"USER=chronos",
@@ -144,7 +136,7 @@ func TestOutcomeMain(t *testing.T) {
 				Tmpfs(fhs.AbsDevShm, 0, 01777).
 
 				// spRuntimeOp
-				Tmpfs(fhs.AbsRunUser, 1<<12, 0755).
+				Tmpfs(fhs.AbsRunUser, xdgRuntimeDirSize, 0755).
 				Bind(m("/tmp/hakurei.0/runtime/9"), m("/run/user/1971"), std.BindWritable).
 
 				// spTmpdirOp
@@ -157,10 +149,6 @@ func TestOutcomeMain(t *testing.T) {
 				// spWaylandOp
 				Bind(m("/tmp/hakurei.0/aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa/wayland"), m("/run/user/1971/wayland-0"), 0).
 
-				// spPulseOp
-				Bind(m("/run/user/1971/hakurei/aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa/pulse"), m("/run/user/1971/pulse/native"), 0).
-				Place(m("/.hakurei/pulse-cookie"), bytes.Repeat([]byte{0}, pulseCookieSizeMax)).
-
 				// spDBusOp
 				Bind(m("/tmp/hakurei.0/aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa/bus"), m("/run/user/1971/bus"), 0).
 				Bind(m("/tmp/hakurei.0/aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa/system_bus_socket"), m("/var/run/dbus/system_bus_socket"), 0).
@@ -182,7 +170,7 @@ func TestOutcomeMain(t *testing.T) {
 				Remount(fhs.AbsRoot, syscall.MS_RDONLY),
 		}},
 
-		{"nixos permissive defaults no enablements", new(stubNixOS), &hst.Config{Container: &hst.ContainerConfig{
+		{"nixos permissive defaults no enablements", new(stubNixOS), &hst.Config{DirectPipeWire: true, Container: &hst.ContainerConfig{
 			Filesystem: []hst.FilesystemConfigJSON{
 				{FilesystemConfig: &hst.FSBind{
 					Target:  fhs.AbsRoot,
@@ -244,7 +232,7 @@ func TestOutcomeMain(t *testing.T) {
 				Tmpfs(hst.AbsPrivateTmp, 4096, 0755).
 				DevWritable(m("/dev/"), true).
 				Tmpfs(m("/dev/shm/"), 0, 01777).
-				Tmpfs(m("/run/user/"), 4096, 0755).
+				Tmpfs(m("/run/user/"), xdgRuntimeDirSize, 0755).
 				Bind(m("/tmp/hakurei.0/runtime/0"), m("/run/user/65534"), std.BindWritable).
 				Bind(m("/tmp/hakurei.0/tmpdir/0"), m("/tmp/"), std.BindWritable).
 				Place(m("/etc/passwd"), []byte("chronos:x:65534:65534:Hakurei:/home/chronos:/run/current-system/sw/bin/zsh\n")).
@@ -264,6 +252,8 @@ func TestOutcomeMain(t *testing.T) {
 		}},
 
 		{"nixos permissive defaults chromium", new(stubNixOS), &hst.Config{
+			DirectPipeWire: true,
+
 			ID:       "org.chromium.Chromium",
 			Identity: 9,
 			Groups:   []string{"video"},
@@ -298,7 +288,7 @@ func TestOutcomeMain(t *testing.T) {
 				},
 				Filter: true,
 			},
-			Enablements: hst.NewEnablements(hst.EWayland | hst.EDBus | hst.EPulse),
+			Enablements: hst.NewEnablements(hst.EWayland | hst.EDBus | hst.EPipeWire | hst.EPulse),
 
 			Container: &hst.ContainerConfig{
 				Filesystem: []hst.FilesystemConfigJSON{
@@ -347,10 +337,7 @@ func TestOutcomeMain(t *testing.T) {
 			Ensure(m("/tmp/hakurei.0/tmpdir/9"), 01700).UpdatePermType(system.User, m("/tmp/hakurei.0/tmpdir/9"), acl.Read, acl.Write, acl.Execute).
 			Ephemeral(system.Process, m("/tmp/hakurei.0/ebf083d1b175911782d413369b64ce7c"), 0711).
 			Wayland(m("/tmp/hakurei.0/ebf083d1b175911782d413369b64ce7c/wayland"), m("/run/user/1971/wayland-0"), "org.chromium.Chromium", "ebf083d1b175911782d413369b64ce7c").
-			Ensure(m("/run/user/1971"), 0700).UpdatePermType(system.User, m("/run/user/1971"), acl.Execute). // this is ordered as is because the previous Ensure only calls mkdir if XDG_RUNTIME_DIR is unset
-			Ensure(m("/run/user/1971/hakurei"), 0700).UpdatePermType(system.User, m("/run/user/1971/hakurei"), acl.Execute).
-			Ephemeral(system.Process, m("/run/user/1971/hakurei/ebf083d1b175911782d413369b64ce7c"), 0700).UpdatePermType(system.Process, m("/run/user/1971/hakurei/ebf083d1b175911782d413369b64ce7c"), acl.Execute).
-			Link(m("/run/user/1971/pulse/native"), m("/run/user/1971/hakurei/ebf083d1b175911782d413369b64ce7c/pulse")).
+			PipeWire(m("/tmp/hakurei.0/ebf083d1b175911782d413369b64ce7c/pipewire"), "org.chromium.Chromium", "ebf083d1b175911782d413369b64ce7c").
 			MustProxyDBus(&hst.BusConfig{
 				Talk: []string{
 					"org.freedesktop.Notifications",
@@ -397,8 +384,7 @@ func TestOutcomeMain(t *testing.T) {
 				"DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/65534/bus",
 				"DBUS_SYSTEM_BUS_ADDRESS=unix:path=/var/run/dbus/system_bus_socket",
 				"HOME=/home/chronos",
-				"PULSE_COOKIE=" + hst.PrivateTmp + "/pulse-cookie",
-				"PULSE_SERVER=unix:/run/user/65534/pulse/native",
+				"PIPEWIRE_REMOTE=/run/user/65534/pipewire-0",
 				"SHELL=/run/current-system/sw/bin/zsh",
 				"TERM=xterm-256color",
 				"USER=chronos",
@@ -413,14 +399,13 @@ func TestOutcomeMain(t *testing.T) {
 				Tmpfs(hst.AbsPrivateTmp, 4096, 0755).
 				DevWritable(m("/dev/"), true).
 				Tmpfs(m("/dev/shm/"), 0, 01777).
-				Tmpfs(m("/run/user/"), 4096, 0755).
+				Tmpfs(m("/run/user/"), xdgRuntimeDirSize, 0755).
 				Bind(m("/tmp/hakurei.0/runtime/9"), m("/run/user/65534"), std.BindWritable).
 				Bind(m("/tmp/hakurei.0/tmpdir/9"), m("/tmp/"), std.BindWritable).
 				Place(m("/etc/passwd"), []byte("chronos:x:65534:65534:Hakurei:/home/chronos:/run/current-system/sw/bin/zsh\n")).
 				Place(m("/etc/group"), []byte("hakurei:x:65534:\n")).
 				Bind(m("/tmp/hakurei.0/ebf083d1b175911782d413369b64ce7c/wayland"), m("/run/user/65534/wayland-0"), 0).
-				Bind(m("/run/user/1971/hakurei/ebf083d1b175911782d413369b64ce7c/pulse"), m("/run/user/65534/pulse/native"), 0).
-				Place(m(hst.PrivateTmp+"/pulse-cookie"), bytes.Repeat([]byte{0}, pulseCookieSizeMax)).
+				Bind(m("/tmp/hakurei.0/ebf083d1b175911782d413369b64ce7c/pipewire"), m("/run/user/65534/pipewire-0"), 0).
 				Bind(m("/tmp/hakurei.0/ebf083d1b175911782d413369b64ce7c/bus"), m("/run/user/65534/bus"), 0).
 				Bind(m("/tmp/hakurei.0/ebf083d1b175911782d413369b64ce7c/system_bus_socket"), m("/var/run/dbus/system_bus_socket"), 0).
 				Bind(m("/dev/dri"), m("/dev/dri"), std.BindWritable|std.BindDevice|std.BindOptional).
@@ -439,8 +424,10 @@ func TestOutcomeMain(t *testing.T) {
 		}},
 
 		{"nixos chromium direct wayland", new(stubNixOS), &hst.Config{
+			DirectPipeWire: true,
+
 			ID:          "org.chromium.Chromium",
-			Enablements: hst.NewEnablements(hst.EWayland | hst.EDBus | hst.EPulse),
+			Enablements: hst.NewEnablements(hst.EWayland | hst.EDBus | hst.EPipeWire | hst.EPulse),
 			Container: &hst.ContainerConfig{
 				Env: nil,
 				Filesystem: []hst.FilesystemConfigJSON{
@@ -502,9 +489,8 @@ func TestOutcomeMain(t *testing.T) {
 			Ensure(m("/run/user/1971"), 0700).UpdatePermType(system.User, m("/run/user/1971"), acl.Execute). // this is ordered as is because the previous Ensure only calls mkdir if XDG_RUNTIME_DIR is unset
 			Ensure(m("/run/user/1971/hakurei"), 0700).UpdatePermType(system.User, m("/run/user/1971/hakurei"), acl.Execute).
 			UpdatePermType(hst.EWayland, m("/run/user/1971/wayland-0"), acl.Read, acl.Write, acl.Execute).
-			Ephemeral(system.Process, m("/run/user/1971/hakurei/8e2c76b066dabe574cf073bdb46eb5c1"), 0700).UpdatePermType(system.Process, m("/run/user/1971/hakurei/8e2c76b066dabe574cf073bdb46eb5c1"), acl.Execute).
-			Link(m("/run/user/1971/pulse/native"), m("/run/user/1971/hakurei/8e2c76b066dabe574cf073bdb46eb5c1/pulse")).
 			Ephemeral(system.Process, m("/tmp/hakurei.0/8e2c76b066dabe574cf073bdb46eb5c1"), 0711).
+			PipeWire(m("/tmp/hakurei.0/8e2c76b066dabe574cf073bdb46eb5c1/pipewire"), "org.chromium.Chromium", "8e2c76b066dabe574cf073bdb46eb5c1").
 			MustProxyDBus(&hst.BusConfig{
 				Talk: []string{
 					"org.freedesktop.FileManager1", "org.freedesktop.Notifications",
@@ -544,8 +530,7 @@ func TestOutcomeMain(t *testing.T) {
 				"DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/1971/bus",
 				"DBUS_SYSTEM_BUS_ADDRESS=unix:path=/var/run/dbus/system_bus_socket",
 				"HOME=/var/lib/persist/module/hakurei/0/1",
-				"PULSE_COOKIE=" + hst.PrivateTmp + "/pulse-cookie",
-				"PULSE_SERVER=unix:/run/user/1971/pulse/native",
+				"PIPEWIRE_REMOTE=/run/user/1971/pipewire-0",
 				"SHELL=/run/current-system/sw/bin/zsh",
 				"TERM=xterm-256color",
 				"USER=u0_a1",
@@ -559,14 +544,13 @@ func TestOutcomeMain(t *testing.T) {
 				Tmpfs(hst.AbsPrivateTmp, 4096, 0755).
 				DevWritable(m("/dev/"), true).
 				Tmpfs(m("/dev/shm/"), 0, 01777).
-				Tmpfs(m("/run/user/"), 4096, 0755).
+				Tmpfs(m("/run/user/"), xdgRuntimeDirSize, 0755).
 				Bind(m("/tmp/hakurei.0/runtime/1"), m("/run/user/1971"), std.BindWritable).
 				Bind(m("/tmp/hakurei.0/tmpdir/1"), m("/tmp/"), std.BindWritable).
 				Place(m("/etc/passwd"), []byte("u0_a1:x:1971:100:Hakurei:/var/lib/persist/module/hakurei/0/1:/run/current-system/sw/bin/zsh\n")).
 				Place(m("/etc/group"), []byte("hakurei:x:100:\n")).
 				Bind(m("/run/user/1971/wayland-0"), m("/run/user/1971/wayland-0"), 0).
-				Bind(m("/run/user/1971/hakurei/8e2c76b066dabe574cf073bdb46eb5c1/pulse"), m("/run/user/1971/pulse/native"), 0).
-				Place(m(hst.PrivateTmp+"/pulse-cookie"), bytes.Repeat([]byte{0}, pulseCookieSizeMax)).
+				Bind(m("/tmp/hakurei.0/8e2c76b066dabe574cf073bdb46eb5c1/pipewire"), m("/run/user/1971/pipewire-0"), 0).
 				Bind(m("/tmp/hakurei.0/8e2c76b066dabe574cf073bdb46eb5c1/bus"), m("/run/user/1971/bus"), 0).
 				Bind(m("/tmp/hakurei.0/8e2c76b066dabe574cf073bdb46eb5c1/system_bus_socket"), m("/var/run/dbus/system_bus_socket"), 0).
 				Bind(m("/bin"), m("/bin"), 0).
@@ -899,6 +883,16 @@ func (k *stubNixOS) lookupGroupId(name string) (string, error) {
 	}
 }
 
+func (k *stubNixOS) lookPath(file string) (string, error) {
+	switch file {
+	case "pipewire-pulse":
+		return "/run/current-system/sw/bin/pipewire-pulse", nil
+
+	default:
+		panic(fmt.Sprintf("unexpected file %q", file))
+	}
+}
+
 func (k *stubNixOS) cmdOutput(cmd *exec.Cmd) ([]byte, error) {
 	switch cmd.Path {
 	case "/proc/nonexistent/hsu":

internal/outcome/shim.go

@@ -14,9 +14,12 @@ import (
 	"time"
 
 	"hakurei.app/container"
+	"hakurei.app/container/check"
+	"hakurei.app/container/fhs"
 	"hakurei.app/container/seccomp"
 	"hakurei.app/container/std"
 	"hakurei.app/hst"
+	"hakurei.app/internal/pipewire"
 	"hakurei.app/message"
 )
 
@@ -83,6 +86,55 @@ func Shim(msg message.Msg) {
 	shimEntrypoint(direct{msg})
 }
 
+// A shimPrivate holds state of the private work directory owned by shim.
+type shimPrivate struct {
+	// Path to directory if created.
+	pathname *check.Absolute
+
+	k  syscallDispatcher
+	id *stringPair[hst.ID]
+}
+
+// unwrap returns the underlying pathname.
+func (sp *shimPrivate) unwrap() *check.Absolute {
+	if sp.pathname == nil {
+		if a, err := check.NewAbs(sp.k.tempdir()); err != nil {
+			sp.k.fatal(err)
+			panic("unreachable")
+		} else {
+			pathname := a.Append(".hakurei-shim-" + sp.id.String())
+			sp.k.getMsg().Verbosef("creating private work directory %q", pathname)
+			if err = sp.k.mkdir(pathname.String(), 0700); err != nil {
+				sp.k.fatal(err)
+				panic("unreachable")
+			}
+			sp.pathname = pathname
+			return sp.unwrap()
+		}
+	} else {
+		return sp.pathname
+	}
+}
+
+// String returns the absolute pathname to the directory held by shimPrivate.
+func (sp *shimPrivate) String() string { return sp.unwrap().String() }
+
+// destroy removes the directory held by shimPrivate.
+func (sp *shimPrivate) destroy() {
+	defer func() { sp.pathname = nil }()
+	if sp.pathname != nil {
+		sp.k.getMsg().Verbosef("destroying private work directory %q", sp.pathname)
+		if err := sp.k.removeAll(sp.pathname.String()); err != nil {
+			sp.k.getMsg().GetLogger().Println(err)
+		}
+	}
+}
+
+const (
+	// shimPipeWireTimeout is the duration pipewire-pulse is allowed to run before its socket becomes available.
+	shimPipeWireTimeout = 5 * time.Second
+)
+
 func shimEntrypoint(k syscallDispatcher) {
 	msg := k.getMsg()
 	if msg == nil {
@@ -208,6 +260,7 @@ func shimEntrypoint(k syscallDispatcher) {
 
 	ctx, stop := k.notifyContext(context.Background(), os.Interrupt, syscall.SIGTERM)
 	cancelContainer.Store(&stop)
+	sp := shimPrivate{k: k, id: state.id}
 	z := container.New(ctx, msg)
 	z.Params = *stateParams.params
 	z.Stdin, z.Stdout, z.Stderr = os.Stdin, os.Stdout, os.Stderr
@@ -215,6 +268,79 @@ func shimEntrypoint(k syscallDispatcher) {
 	// bounds and default enforced in finalise.go
 	z.WaitDelay = state.Shim.WaitDelay
 
+	if stateParams.pipewirePulsePath != nil {
+		zpw := container.NewCommand(ctx, msg, stateParams.pipewirePulsePath, pipewirePulseName)
+		zpw.Hostname = "hakurei-" + pipewirePulseName
+		zpw.SeccompFlags |= seccomp.AllowMultiarch
+		zpw.SeccompPresets |= std.PresetStrict
+		zpw.Env = []string{
+			// pipewire SecurityContext socket path
+			pipewire.Remote + "=" + stateParams.instancePath().Append("pipewire").String(),
+			// pipewire-pulse socket directory path
+			envXDGRuntimeDir + "=" + sp.String(),
+		}
+		if msg.IsVerbose() {
+			zpw.Stdin, zpw.Stdout, zpw.Stderr = os.Stdin, os.Stdout, os.Stderr
+		}
+		zpw.
+			Bind(fhs.AbsRoot, fhs.AbsRoot, 0).
+			Bind(sp.unwrap(), sp.unwrap(), std.BindWritable).
+			Proc(fhs.AbsProc).Dev(fhs.AbsDev, true)
+		socketPath := sp.unwrap().Append("pulse", "native")
+		innerSocketPath := stateParams.runtimeDir.Append("pulse", "native")
+
+		if err := k.containerStart(zpw); err != nil {
+			sp.destroy()
+			printMessageError(func(v ...any) { k.fatal(fmt.Sprintln(v...)) },
+				"cannot start "+pipewirePulseName+" container:", err)
+		}
+		if err := k.containerServe(zpw); err != nil {
+			sp.destroy()
+			printMessageError(func(v ...any) { k.fatal(fmt.Sprintln(v...)) },
+				"cannot configure "+pipewirePulseName+" container:", err)
+		}
+
+		done := make(chan error, 1)
+		k.new(func(k syscallDispatcher, msg message.Msg) { done <- k.containerWait(zpw) })
+
+		socketTimer := time.NewTimer(shimPipeWireTimeout)
+		for {
+			select {
+			case <-socketTimer.C:
+				sp.destroy()
+				k.fatal(pipewirePulseName + " exceeded deadline before socket appeared")
+				break
+
+			case err := <-done:
+				var exitError *exec.ExitError
+				if !errors.As(err, &exitError) {
+					msg.Verbosef("cannot wait: %v", err)
+					k.exit(127)
+				}
+				sp.destroy()
+				k.fatal(pipewirePulseName + " " + exitError.ProcessState.String())
+				break
+
+			default:
+				if _, err := k.stat(socketPath.String()); err != nil {
+					if !errors.Is(err, os.ErrNotExist) {
+						sp.destroy()
+						k.fatal(err)
+						break
+					}
+
+					time.Sleep(500 * time.Microsecond)
+					continue
+				}
+			}
+
+			break
+		}
+
+		z.Bind(socketPath, innerSocketPath, 0)
+		z.Env = append(z.Env, "PULSE_SERVER=unix:"+innerSocketPath.String())
+	}
+
 	if err := k.containerStart(z); err != nil {
 		var f func(v ...any)
 		if logger := msg.GetLogger(); logger != nil {
@@ -225,9 +351,11 @@ func shimEntrypoint(k syscallDispatcher) {
 			}
 		}
 		printMessageError(f, "cannot start container:", err)
+		sp.destroy()
 		k.exit(hst.ExitFailure)
 	}
 	if err := k.containerServe(z); err != nil {
+		sp.destroy()
 		printMessageError(func(v ...any) { k.fatal(fmt.Sprintln(v...)) },
 			"cannot configure container:", err)
 	}
@@ -236,10 +364,13 @@ func shimEntrypoint(k syscallDispatcher) {
 		seccomp.Preset(std.PresetStrict, seccomp.AllowMultiarch),
 		seccomp.AllowMultiarch,
 	); err != nil {
+		sp.destroy()
 		k.fatalf("cannot load syscall filter: %v", err)
 	}
 
 	if err := k.containerWait(z); err != nil {
+		sp.destroy()
+
 		var exitError *exec.ExitError
 		if !errors.As(err, &exitError) {
 			if errors.Is(err, context.Canceled) {
@@ -250,4 +381,5 @@ func shimEntrypoint(k syscallDispatcher) {
 		}
 		k.exit(exitError.ExitCode())
 	}
+	sp.destroy()
 }

internal/outcome/shim_test.go

@@ -69,7 +69,7 @@ func TestShimEntrypoint(t *testing.T) {
 			Tmpfs(fhs.AbsDevShm, 0, 01777).
 
 			// spRuntimeOp
-			Tmpfs(fhs.AbsRunUser, 1<<12, 0755).
+			Tmpfs(fhs.AbsRunUser, xdgRuntimeDirSize, 0755).
 			Bind(m("/tmp/hakurei.10/runtime/9999"), m("/run/user/1000"), std.BindWritable).
 
 			// spTmpdirOp

internal/outcome/spcontainer.go

@@ -382,6 +382,10 @@ func (p opsAdapter) Link(target *check.Absolute, linkName string, dereference bo
 	return opsAdapter{p.Ops.Link(target, linkName, dereference)}
 }
 
+func (p opsAdapter) Daemon(target, path *check.Absolute, args ...string) hst.Ops {
+	return opsAdapter{p.Ops.Daemon(target, path, args...)}
+}
+
 func (p opsAdapter) Root(host *check.Absolute, flags int) hst.Ops {
 	return opsAdapter{p.Ops.Root(host, flags)}
 }

internal/outcome/sppipewire.go

@@ -0,0 +1,53 @@
+package outcome
+
+import (
+	"encoding/gob"
+
+	"hakurei.app/container/check"
+	"hakurei.app/hst"
+	"hakurei.app/internal/pipewire"
+)
+
+const pipewirePulseName = "pipewire-pulse"
+
+func init() { gob.Register(new(spPipeWireOp)) }
+
+// spPipeWireOp exports the PipeWire server to the container via SecurityContext.
+// Runs after spRuntimeOp.
+type spPipeWireOp struct {
+	// Path to pipewire-pulse server. Populated during toSystem if DirectPipeWire is false.
+	CompatServerPath *check.Absolute
+}
+
+func (s *spPipeWireOp) toSystem(state *outcomeStateSys) error {
+	if state.et&hst.EPipeWire == 0 {
+		return errNotEnabled
+	}
+	if !state.directPipeWire {
+		if n, err := state.k.lookPath(pipewirePulseName); err != nil {
+			return &hst.AppError{Step: "look up " + pipewirePulseName, Err: err}
+		} else if s.CompatServerPath, err = check.NewAbs(n); err != nil {
+			return err
+		}
+	}
+
+	appId := state.appId
+	if appId == "" {
+		// use instance ID in case app id is not set
+		appId = "app.hakurei." + state.id.String()
+	}
+	state.sys.PipeWire(state.instance().Append("pipewire"), appId, state.id.String())
+	return nil
+}
+
+func (s *spPipeWireOp) toContainer(state *outcomeStateParams) error {
+	if s.CompatServerPath == nil {
+		innerPath := state.runtimeDir.Append(pipewire.PW_DEFAULT_REMOTE)
+		state.env[pipewire.Remote] = innerPath.String()
+		state.params.Bind(state.instancePath().Append("pipewire"), innerPath, 0)
+	}
+
+	// pipewire-pulse behaviour implemented in shim.go
+	state.pipewirePulsePath = s.CompatServerPath
+	return nil
+}

internal/outcome/sppipewire_test.go

@@ -0,0 +1,49 @@
+package outcome
+
+import (
+	"testing"
+
+	"hakurei.app/container"
+	"hakurei.app/container/stub"
+	"hakurei.app/hst"
+	"hakurei.app/internal/pipewire"
+	"hakurei.app/internal/system"
+)
+
+func TestSpPipeWireOp(t *testing.T) {
+	t.Parallel()
+	config := hst.Template()
+
+	checkOpBehaviour(t, []opBehaviourTestCase{
+		{"not enabled", func(bool, bool) outcomeOp {
+			return new(spPipeWireOp)
+		}, func() *hst.Config {
+			c := hst.Template()
+			*c.Enablements = 0
+			return c
+		}, nil, nil, nil, nil, errNotEnabled, nil, nil, nil, nil, nil},
+
+		{"success", func(bool, bool) outcomeOp {
+			return new(spPipeWireOp)
+		}, func() *hst.Config {
+			c := hst.Template()
+			c.DirectPipeWire = true
+			return c
+		}, nil, []stub.Call{}, newI().
+			// state.instance
+			Ephemeral(system.Process, m(wantInstancePrefix), 0711).
+			// toSystem
+			PipeWire(
+				m(wantInstancePrefix+"/pipewire"),
+				"org.chromium.Chromium",
+				"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa",
+			), sysUsesInstance(nil), nil, insertsOps(afterSpRuntimeOp(nil)), []stub.Call{
+			// this op configures the container state and does not make calls during toContainer
+		}, &container.Params{
+			Ops: new(container.Ops).
+				Bind(m(wantInstancePrefix+"/pipewire"), m("/run/user/1000/pipewire-0"), 0),
+		}, paramsWantEnv(config, map[string]string{
+			pipewire.Remote: "/run/user/1000/pipewire-0",
+		}, nil), nil},
+	})
+}

internal/outcome/sppulse.go

@@ -29,7 +29,7 @@ type spPulseOp struct {
 }
 
 func (s *spPulseOp) toSystem(state *outcomeStateSys) error {
-	if state.et&hst.EPulse == 0 {
+	if !state.directPulse || state.et&hst.EPulse == 0 {
 		return errNotEnabled
 	}
 

internal/outcome/sppulse_test.go

@@ -18,24 +18,40 @@ import (
 func TestSpPulseOp(t *testing.T) {
 	t.Parallel()
 
-	config := hst.Template()
+	newConfig := func() *hst.Config {
+		config := hst.Template()
+		config.DirectPulse = true
+		config.Enablements = hst.NewEnablements(hst.EPulse)
+		return config
+	}
+
+	config := newConfig()
 	sampleCookie := bytes.Repeat([]byte{0xfc}, pulseCookieSizeMax)
 
 	checkOpBehaviour(t, []opBehaviourTestCase{
 		{"not enabled", func(bool, bool) outcomeOp {
 			return new(spPulseOp)
 		}, func() *hst.Config {
-			c := hst.Template()
+			c := newConfig()
+			c.DirectPulse = true
 			*c.Enablements = 0
 			return c
 		}, nil, nil, nil, nil, errNotEnabled, nil, nil, nil, nil, nil},
 
+		{"not enabled direct", func(bool, bool) outcomeOp {
+			return new(spPulseOp)
+		}, func() *hst.Config {
+			c := newConfig()
+			c.DirectPulse = false
+			return c
+		}, nil, nil, nil, nil, errNotEnabled, nil, nil, nil, nil, nil},
+
 		{"socketDir stat", func(isShim, _ bool) outcomeOp {
 			if !isShim {
 				return new(spPulseOp)
 			}
 			return &spPulseOp{Cookie: (*[256]byte)(sampleCookie)}
-		}, hst.Template, nil, []stub.Call{
+		}, newConfig, nil, []stub.Call{
 			call("stat", stub.ExpectArgs{wantRuntimePath + "/pulse"}, (*stubFi)(nil), stub.UniqueError(2)),
 		}, nil, nil, &hst.AppError{
 			Step: `access PulseAudio directory "/proc/nonexistent/xdg_runtime_dir/pulse"`,
@@ -44,7 +60,7 @@ func TestSpPulseOp(t *testing.T) {
 
 		{"socketDir nonexistent", func(bool, bool) outcomeOp {
 			return new(spPulseOp)
-		}, hst.Template, nil, []stub.Call{
+		}, newConfig, nil, []stub.Call{
 			call("stat", stub.ExpectArgs{wantRuntimePath + "/pulse"}, (*stubFi)(nil), os.ErrNotExist),
 		}, nil, nil, &hst.AppError{
 			Step: "finalise",
@@ -54,7 +70,7 @@ func TestSpPulseOp(t *testing.T) {
 
 		{"socket stat", func(bool, bool) outcomeOp {
 			return new(spPulseOp)
-		}, hst.Template, nil, []stub.Call{
+		}, newConfig, nil, []stub.Call{
 			call("stat", stub.ExpectArgs{wantRuntimePath + "/pulse"}, (*stubFi)(nil), nil),
 			call("stat", stub.ExpectArgs{wantRuntimePath + "/pulse/native"}, (*stubFi)(nil), stub.UniqueError(1)),
 		}, nil, nil, &hst.AppError{
@@ -64,7 +80,7 @@ func TestSpPulseOp(t *testing.T) {
 
 		{"socket nonexistent", func(bool, bool) outcomeOp {
 			return new(spPulseOp)
-		}, hst.Template, nil, []stub.Call{
+		}, newConfig, nil, []stub.Call{
 			call("stat", stub.ExpectArgs{wantRuntimePath + "/pulse"}, (*stubFi)(nil), nil),
 			call("stat", stub.ExpectArgs{wantRuntimePath + "/pulse/native"}, (*stubFi)(nil), os.ErrNotExist),
 		}, nil, nil, &hst.AppError{
@@ -75,7 +91,7 @@ func TestSpPulseOp(t *testing.T) {
 
 		{"socket mode", func(bool, bool) outcomeOp {
 			return new(spPulseOp)
-		}, hst.Template, nil, []stub.Call{
+		}, newConfig, nil, []stub.Call{
 			call("stat", stub.ExpectArgs{wantRuntimePath + "/pulse"}, (*stubFi)(nil), nil),
 			call("stat", stub.ExpectArgs{wantRuntimePath + "/pulse/native"}, &stubFi{mode: 0660}, nil),
 		}, nil, nil, &hst.AppError{
@@ -86,18 +102,18 @@ func TestSpPulseOp(t *testing.T) {
 
 		{"cookie notAbs", func(bool, bool) outcomeOp {
 			return new(spPulseOp)
-		}, hst.Template, nil, []stub.Call{
+		}, newConfig, nil, []stub.Call{
 			call("stat", stub.ExpectArgs{wantRuntimePath + "/pulse"}, (*stubFi)(nil), nil),
 			call("stat", stub.ExpectArgs{wantRuntimePath + "/pulse/native"}, &stubFi{mode: 0666}, nil),
 			call("lookupEnv", stub.ExpectArgs{"PULSE_COOKIE"}, "proc/nonexistent/cookie", nil),
 		}, nil, nil, &hst.AppError{
 			Step: "locate PulseAudio cookie",
-			Err:  &check.AbsoluteError{Pathname: "proc/nonexistent/cookie"},
+			Err:  check.AbsoluteError("proc/nonexistent/cookie"),
 		}, nil, nil, nil, nil, nil},
 
 		{"cookie loadFile", func(bool, bool) outcomeOp {
 			return new(spPulseOp)
-		}, hst.Template, nil, []stub.Call{
+		}, newConfig, nil, []stub.Call{
 			call("stat", stub.ExpectArgs{wantRuntimePath + "/pulse"}, (*stubFi)(nil), nil),
 			call("stat", stub.ExpectArgs{wantRuntimePath + "/pulse/native"}, &stubFi{mode: 0666}, nil),
 			call("lookupEnv", stub.ExpectArgs{"PULSE_COOKIE"}, "/proc/nonexistent/cookie", nil),
@@ -118,7 +134,7 @@ func TestSpPulseOp(t *testing.T) {
 				op.CookieSize += +0xfd
 			}
 			return op
-		}, hst.Template, nil, []stub.Call{
+		}, newConfig, nil, []stub.Call{
 			call("stat", stub.ExpectArgs{wantRuntimePath + "/pulse"}, (*stubFi)(nil), nil),
 			call("stat", stub.ExpectArgs{wantRuntimePath + "/pulse/native"}, &stubFi{mode: 0666}, nil),
 			call("lookupEnv", stub.ExpectArgs{"PULSE_COOKIE"}, "/proc/nonexistent/cookie", nil),
@@ -150,7 +166,7 @@ func TestSpPulseOp(t *testing.T) {
 			sampleCookieTrunc := make([]byte, pulseCookieSizeMax)
 			copy(sampleCookieTrunc, sampleCookie[:len(sampleCookie)-0xe])
 			return &spPulseOp{Cookie: (*[pulseCookieSizeMax]byte)(sampleCookieTrunc), CookieSize: pulseCookieSizeMax - 0xe}
-		}, hst.Template, nil, []stub.Call{
+		}, newConfig, nil, []stub.Call{
 			call("stat", stub.ExpectArgs{wantRuntimePath + "/pulse"}, (*stubFi)(nil), nil),
 			call("stat", stub.ExpectArgs{wantRuntimePath + "/pulse/native"}, &stubFi{mode: 0666}, nil),
 			call("lookupEnv", stub.ExpectArgs{"PULSE_COOKIE"}, "/proc/nonexistent/cookie", nil),
@@ -183,7 +199,7 @@ func TestSpPulseOp(t *testing.T) {
 				return new(spPulseOp)
 			}
 			return &spPulseOp{Cookie: (*[pulseCookieSizeMax]byte)(sampleCookie), CookieSize: pulseCookieSizeMax}
-		}, hst.Template, nil, []stub.Call{
+		}, newConfig, nil, []stub.Call{
 			call("stat", stub.ExpectArgs{wantRuntimePath + "/pulse"}, (*stubFi)(nil), nil),
 			call("stat", stub.ExpectArgs{wantRuntimePath + "/pulse/native"}, &stubFi{mode: 0666}, nil),
 			call("lookupEnv", stub.ExpectArgs{"PULSE_COOKIE"}, "/proc/nonexistent/cookie", nil),
@@ -213,7 +229,7 @@ func TestSpPulseOp(t *testing.T) {
 
 		{"success", func(bool, bool) outcomeOp {
 			return new(spPulseOp)
-		}, hst.Template, nil, []stub.Call{
+		}, newConfig, nil, []stub.Call{
 			call("stat", stub.ExpectArgs{wantRuntimePath + "/pulse"}, (*stubFi)(nil), nil),
 			call("stat", stub.ExpectArgs{wantRuntimePath + "/pulse/native"}, &stubFi{mode: 0666}, nil),
 			call("lookupEnv", stub.ExpectArgs{"PULSE_COOKIE"}, nil, nil),
@@ -256,7 +272,7 @@ func TestDiscoverPulseCookie(t *testing.T) {
 			call("verbose", stub.ExpectArgs{[]any{(*check.Absolute)(nil)}}, nil, nil),
 		}}, &hst.AppError{
 			Step: "locate PulseAudio cookie",
-			Err:  &check.AbsoluteError{Pathname: "proc/nonexistent/pulse-cookie"},
+			Err:  check.AbsoluteError("proc/nonexistent/pulse-cookie"),
 		}},
 
 		{"success override", fCheckPathname, stub.Expect{Calls: []stub.Call{
@@ -270,7 +286,7 @@ func TestDiscoverPulseCookie(t *testing.T) {
 			call("verbose", stub.ExpectArgs{[]any{(*check.Absolute)(nil)}}, nil, nil),
 		}}, &hst.AppError{
 			Step: "locate PulseAudio cookie",
-			Err:  &check.AbsoluteError{Pathname: "proc/nonexistent/home"},
+			Err:  check.AbsoluteError("proc/nonexistent/home"),
 		}},
 
 		{"home stat", fCheckPathname, stub.Expect{Calls: []stub.Call{
@@ -305,7 +321,7 @@ func TestDiscoverPulseCookie(t *testing.T) {
 			call("verbose", stub.ExpectArgs{[]any{(*check.Absolute)(nil)}}, nil, nil),
 		}}, &hst.AppError{
 			Step: "locate PulseAudio cookie",
-			Err:  &check.AbsoluteError{Pathname: "proc/nonexistent/xdg"},
+			Err:  check.AbsoluteError("proc/nonexistent/xdg"),
 		}},
 
 		{"xdg stat", fCheckPathname, stub.Expect{Calls: []stub.Call{

internal/outcome/spruntime.go

@@ -91,6 +91,9 @@ func (s *spRuntimeOp) toSystem(state *outcomeStateSys) error {
 	return nil
 }
 
+// xdgRuntimeDirSize is the size of the filesystem mounted on inner XDG_RUNTIME_DIR.
+const xdgRuntimeDirSize = 1 << 24
+
 func (s *spRuntimeOp) toContainer(state *outcomeStateParams) error {
 	state.runtimeDir = fhs.AbsRunUser.Append(state.mapuid.String())
 	state.env[envXDGRuntimeDir] = state.runtimeDir.String()
@@ -108,7 +111,7 @@ func (s *spRuntimeOp) toContainer(state *outcomeStateParams) error {
 
 	}
 
-	state.params.Tmpfs(fhs.AbsRunUser, 1<<12, 0755)
+	state.params.Tmpfs(fhs.AbsRunUser, xdgRuntimeDirSize, 0755)
 	if state.Container.Flags&hst.FShareRuntime != 0 {
 		_, runtimeDirInst := s.commonPaths(state.outcomeState)
 		state.params.Bind(runtimeDirInst, state.runtimeDir, std.BindWritable)

internal/outcome/spruntime_test.go

@@ -40,7 +40,7 @@ func TestSpRuntimeOp(t *testing.T) {
 			// this op configures the container state and does not make calls during toContainer
 		}, &container.Params{
 			Ops: new(container.Ops).
-				Tmpfs(fhs.AbsRunUser, 1<<12, 0755).
+				Tmpfs(fhs.AbsRunUser, xdgRuntimeDirSize, 0755).
 				Bind(m("/proc/nonexistent/tmp/hakurei.0/runtime/9"), m("/run/user/1000"), std.BindWritable),
 		}, paramsWantEnv(config, map[string]string{
 			"XDG_RUNTIME_DIR":   "/run/user/1000",
@@ -67,7 +67,7 @@ func TestSpRuntimeOp(t *testing.T) {
 			// this op configures the container state and does not make calls during toContainer
 		}, &container.Params{
 			Ops: new(container.Ops).
-				Tmpfs(fhs.AbsRunUser, 1<<12, 0755).
+				Tmpfs(fhs.AbsRunUser, xdgRuntimeDirSize, 0755).
 				Bind(m("/proc/nonexistent/tmp/hakurei.0/runtime/9"), m("/run/user/1000"), std.BindWritable),
 		}, paramsWantEnv(config, map[string]string{
 			"XDG_RUNTIME_DIR":   "/run/user/1000",
@@ -94,7 +94,7 @@ func TestSpRuntimeOp(t *testing.T) {
 			// this op configures the container state and does not make calls during toContainer
 		}, &container.Params{
 			Ops: new(container.Ops).
-				Tmpfs(fhs.AbsRunUser, 1<<12, 0755).
+				Tmpfs(fhs.AbsRunUser, xdgRuntimeDirSize, 0755).
 				Bind(m("/proc/nonexistent/tmp/hakurei.0/runtime/9"), m("/run/user/1000"), std.BindWritable),
 		}, paramsWantEnv(config, map[string]string{
 			"XDG_RUNTIME_DIR":   "/run/user/1000",
@@ -117,7 +117,7 @@ func TestSpRuntimeOp(t *testing.T) {
 			// this op configures the container state and does not make calls during toContainer
 		}, &container.Params{
 			Ops: new(container.Ops).
-				Tmpfs(fhs.AbsRunUser, 1<<12, 0755).
+				Tmpfs(fhs.AbsRunUser, xdgRuntimeDirSize, 0755).
 				Bind(m("/proc/nonexistent/tmp/hakurei.0/runtime/9"), m("/run/user/1000"), std.BindWritable),
 		}, paramsWantEnv(config, map[string]string{
 			"XDG_RUNTIME_DIR":   "/run/user/1000",

internal/pipewire/client.go

@@ -103,6 +103,8 @@ type Client struct {
 
 	// Populated by [CoreBoundProps] events targeting [Client].
 	Properties SPADict `json:"props"`
+
+	noRemove
 }
 
 func (client *Client) consume(opcode byte, files []int, unmarshal func(v any)) error {
@@ -113,7 +115,7 @@ func (client *Client) consume(opcode byte, files []int, unmarshal func(v any)) e
 		return nil
 
 	default:
-		return &UnsupportedOpcodeError{opcode, client.String()}
+		panic(&UnsupportedOpcodeError{opcode, client.String()})
 	}
 }
 

internal/pipewire/core.go

@@ -40,13 +40,14 @@ const (
 	PW_CORE_EVENT_ADD_MEM
 	PW_CORE_EVENT_REMOVE_MEM
 	PW_CORE_EVENT_BOUND_PROPS
-	PW_CORE_EVENT_NUM
 
+	PW_CORE_EVENT_NUM
 	PW_VERSION_CORE_EVENTS = 1
 )
 
 const (
 	PW_CORE_METHOD_ADD_LISTENER = iota
+
 	PW_CORE_METHOD_HELLO
 	PW_CORE_METHOD_SYNC
 	PW_CORE_METHOD_PONG
@@ -54,25 +55,26 @@ const (
 	PW_CORE_METHOD_GET_REGISTRY
 	PW_CORE_METHOD_CREATE_OBJECT
 	PW_CORE_METHOD_DESTROY
-	PW_CORE_METHOD_NUM
 
+	PW_CORE_METHOD_NUM
 	PW_VERSION_CORE_METHODS = 0
 )
 
 const (
 	PW_REGISTRY_EVENT_GLOBAL = iota
 	PW_REGISTRY_EVENT_GLOBAL_REMOVE
-	PW_REGISTRY_EVENT_NUM
 
+	PW_REGISTRY_EVENT_NUM
 	PW_VERSION_REGISTRY_EVENTS = 0
 )
 
 const (
 	PW_REGISTRY_METHOD_ADD_LISTENER = iota
+
 	PW_REGISTRY_METHOD_BIND
 	PW_REGISTRY_METHOD_DESTROY
-	PW_REGISTRY_METHOD_NUM
 
+	PW_REGISTRY_METHOD_NUM
 	PW_VERSION_REGISTRY_METHODS = 0
 )
 
@@ -266,6 +268,31 @@ type CoreErrorEvent struct{ CoreError }
 // Opcode satisfies [Message] with a constant value.
 func (c *CoreErrorEvent) Opcode() byte { return PW_CORE_EVENT_ERROR }
 
+// The CoreRemoveId event is used internally by the object ID management logic.
+//
+// When a client deletes an object, the server will send this event to acknowledge
+// that it has seen the delete request. When the client receives this event, it
+// will know that it can safely reuse the object ID.
+type CoreRemoveId struct {
+	// A proxy id that was removed.
+	ID Int `json:"id"`
+}
+
+// Opcode satisfies [Message] with a constant value.
+func (c *CoreRemoveId) Opcode() byte { return PW_CORE_EVENT_REMOVE_ID }
+
+// FileCount satisfies [Message] with a constant value.
+func (c *CoreRemoveId) FileCount() Int { return 0 }
+
+// Size satisfies [KnownSize] with a constant value.
+func (c *CoreRemoveId) Size() Word { return SizePrefix + Size(SizeInt) }
+
+// MarshalBinary satisfies [encoding.BinaryMarshaler] via [Marshal].
+func (c *CoreRemoveId) MarshalBinary() ([]byte, error) { return Marshal(c) }
+
+// UnmarshalBinary satisfies [encoding.BinaryUnmarshaler] via [Unmarshal].
+func (c *CoreRemoveId) UnmarshalBinary(data []byte) error { return Unmarshal(data, c) }
+
 // The CoreBoundProps event is emitted when a local object ID is bound to a global ID.
 // It is emitted before the global becomes visible in the registry.
 type CoreBoundProps struct {
@@ -299,13 +326,96 @@ func (c *CoreBoundProps) UnmarshalBinary(data []byte) error { return Unmarshal(d
 
 // ErrBadBoundProps is returned when a [CoreBoundProps] event targeting a proxy
 // that should never be targeted is received and processed.
-var ErrBadBoundProps = errors.New("attempted to store bound props on proxy that should never be targeted")
+var ErrBadBoundProps = errors.New("attempting to store bound props on a proxy that should never be targeted")
 
 // noAck is embedded by proxies that are never targeted by [CoreBoundProps].
 type noAck struct{}
 
 // setBoundProps should never be called as this proxy should never be targeted by [CoreBoundProps].
-func (noAck) setBoundProps(*CoreBoundProps) error { return ErrBadBoundProps }
+func (noAck) setBoundProps(*CoreBoundProps) error { panic(ErrBadBoundProps) }
+
+// ErrBadRemove is returned when a [CoreRemoveId] event targeting a proxy
+// that should never be targeted is received and processed.
+var ErrBadRemove = errors.New("attempting to remove a proxy that should never be targeted")
+
+// noRemove is embedded by proxies that are never targeted by [CoreRemoveId].
+type noRemove struct{}
+
+// remove should never be called as this proxy should never be targeted by [CoreRemoveId].
+func (noRemove) remove() error { panic(ErrBadRemove) }
+
+// ErrInvalidRemove is returned when a proxy is somehow removed twice. This is only reached for
+// an implementation error as the proxy struct should no longer be reachable after the first call.
+var ErrInvalidRemove = errors.New("attempting to remove an already freed proxy")
+
+// removable is embedded by proxies that can be targeted by [CoreRemoveId] and requires no cleanup.
+type removable bool
+
+// remove checks against removal of a freed proxy and marks the proxy as removed.
+func (s *removable) remove() error {
+	if *s {
+		panic(ErrInvalidRemove)
+	}
+	*s = true
+	return nil
+}
+
+// ErrProxyDestroyed is returned when attempting to use a proxy method when the underlying
+// proxy has already been targeted by a [CoreRemoveId] event.
+var ErrProxyDestroyed = errors.New("underlying proxy has been removed")
+
+// checkDestroy returns [ErrProxyDestroyed] if the current proxy has been destroyed.
+// Must be called at the beginning of any exported method of a proxy embedding removable.
+func (s *removable) checkDestroy() error {
+	if *s {
+		// not fatal: the caller is allowed to recover from this and allocate a new proxy
+		return ErrProxyDestroyed
+	}
+	return nil
+}
+
+// mustCheckDestroy calls checkDestroy and panics if a non-nil error is returned.
+// This is useful for non-exported methods as they should become unreachable.
+func (s *removable) mustCheckDestroy() {
+	if err := s.checkDestroy(); err != nil {
+		panic(err)
+	}
+}
+
+// destructible is embedded by proxies that can be targeted by the [CoreRemoveId] event and the
+// [CoreDestroy] method and requires no cleanup. destructible purposefully does not override
+// removable.mustCheckDestroy because it is used by unexported methods called during event handling
+// and are exempt from the destruction check.
+type destructible struct {
+	destroyed bool
+
+	removable
+}
+
+// checkDestroy overrides removable.checkDestroy to also check the destroyed field.
+func (s *destructible) checkDestroy() error {
+	if s.destroyed {
+		return ErrProxyDestroyed
+	}
+	if err := s.removable.checkDestroy(); err != nil {
+		return err
+	}
+	return nil
+}
+
+// destroy calls removable.checkDestroy then queues a [CoreDestroy] event if it succeeds.
+func (s *destructible) destroy(ctx *Context, id Int) error {
+	if err := s.checkDestroy(); err != nil {
+		return err
+	}
+	l := len(ctx.pendingDestruction)
+	ctx.pendingDestruction[id] = struct{}{}
+	if len(ctx.pendingDestruction) != l+1 {
+		return ErrProxyDestroyed
+	}
+	s.destroyed = true
+	return ctx.GetCore().destroy(id)
+}
 
 // An InconsistentIdError describes an inconsistent state where the server claims an impossible
 // proxy or global id. This is only generated by the [CoreBoundProps] event.
@@ -349,10 +459,10 @@ func (c *CoreHello) MarshalBinary() ([]byte, error) { return Marshal(c) }
 // UnmarshalBinary satisfies [encoding.BinaryUnmarshaler] via [Unmarshal].
 func (c *CoreHello) UnmarshalBinary(data []byte) error { return Unmarshal(data, c) }
 
-// coreHello queues a [CoreHello] message for the PipeWire server.
-// This method should not be called directly, the New function queues this message.
-func (ctx *Context) coreHello() error {
-	return ctx.writeMessage(
+// hello queues a [CoreHello] message for the PipeWire server.
+// This method should not be called directly, the [New] function queues this message.
+func (core *Core) hello() error {
+	return core.ctx.writeMessage(
 		PW_ID_CORE,
 		&CoreHello{PW_VERSION_CORE},
 	)
@@ -388,12 +498,12 @@ func (c *CoreSync) MarshalBinary() ([]byte, error) { return Marshal(c) }
 // UnmarshalBinary satisfies [encoding.BinaryUnmarshaler] via [Unmarshal].
 func (c *CoreSync) UnmarshalBinary(data []byte) error { return Unmarshal(data, c) }
 
-// coreSync queues a [CoreSync] message for the PipeWire server.
+// sync queues a [CoreSync] message for the PipeWire server.
 // This is not safe to use directly, callers should use Sync instead.
-func (ctx *Context) coreSync(id Int) error {
-	return ctx.writeMessage(
+func (core *Core) sync(id Int) error {
+	return core.ctx.writeMessage(
 		PW_ID_CORE,
-		&CoreSync{id, CoreSyncSequenceOffset + Int(ctx.sequence)},
+		&CoreSync{id, CoreSyncSequenceOffset + Int(core.ctx.sequence)},
 	)
 }
 
@@ -404,13 +514,13 @@ var ErrNotDone = errors.New("did not receive a Core::Done event targeting previo
 const (
 	// syncTimeout is the maximum duration [Core.Sync] is allowed to take before
 	// receiving [CoreDone] or failing.
-	syncTimeout = 5 * time.Second
+	syncTimeout = 10 * time.Second
 )
 
 // Sync queues a [CoreSync] message for the PipeWire server and initiates a Roundtrip.
 func (core *Core) Sync() error {
 	core.done = false
-	if err := core.ctx.coreSync(roundtripSyncID); err != nil {
+	if err := core.sync(roundtripSyncID); err != nil {
 		return err
 	}
 	deadline := time.Now().Add(syncTimeout)
@@ -429,6 +539,10 @@ func (core *Core) Sync() error {
 		core.ctx.closeReceivedFiles()
 		return &ProxyFatalError{Err: UnacknowledgedProxyError(slices.Collect(maps.Keys(core.ctx.pendingIds))), ProxyErrs: core.ctx.cloneAsProxyErrors()}
 	}
+	if len(core.ctx.pendingDestruction) != 0 {
+		core.ctx.closeReceivedFiles()
+		return &ProxyFatalError{Err: UnacknowledgedProxyDestructionError(slices.Collect(maps.Keys(core.ctx.pendingDestruction))), ProxyErrs: core.ctx.cloneAsProxyErrors()}
+	}
 	return core.ctx.doSyncComplete()
 }
 
@@ -497,6 +611,89 @@ func (ctx *Context) GetRegistry() (*Registry, error) {
 	)
 }
 
+// CoreCreateObject is sent when the client requests to create a
+// new object from a factory of a certain type.
+//
+// The client allocates a new_id for the proxy. The server will
+// allocate a new resource with the same new_id and from then on,
+// Methods and Events will be exchanged between the new object of
+// the given type.
+type CoreCreateObject struct {
+	// The name of a server factory object to use.
+	FactoryName String `json:"factory_name"`
+	// The type of the object to create, this is also the type of
+	// the interface of the new_id proxy.
+	Type String `json:"type"`
+	// Undocumented, assumed to be the local version of the proxy.
+	Version Int `json:"version"`
+	// Extra properties to create the object.
+	Properties *SPADict `json:"props"`
+	// The proxy id of the new object.
+	NewID Int `json:"new_id"`
+}
+
+// Opcode satisfies [Message] with a constant value.
+func (c *CoreCreateObject) Opcode() byte { return PW_CORE_METHOD_CREATE_OBJECT }
+
+// FileCount satisfies [Message] with a constant value.
+func (c *CoreCreateObject) FileCount() Int { return 0 }
+
+// Size satisfies [KnownSize] with a value computed at runtime.
+func (c *CoreCreateObject) Size() Word {
+	return SizePrefix +
+		SizeString[Word](c.FactoryName) +
+		SizeString[Word](c.Type) +
+		Size(SizeInt) +
+		c.Properties.Size() +
+		Size(SizeInt)
+}
+
+// MarshalBinary satisfies [encoding.BinaryMarshaler] via [Marshal].
+func (c *CoreCreateObject) MarshalBinary() ([]byte, error) { return Marshal(c) }
+
+// UnmarshalBinary satisfies [encoding.BinaryUnmarshaler] via [Unmarshal].
+func (c *CoreCreateObject) UnmarshalBinary(data []byte) error { return Unmarshal(data, c) }
+
+// createObject queues a [CoreCreateObject] message for the PipeWire server.
+// This is not safe to use directly, callers should use typed wrapper methods on [Registry] instead.
+func (core *Core) createObject(factoryName, typeName String, version Int, props SPADict, newId Int) error {
+	return core.ctx.writeMessage(
+		PW_ID_CORE,
+		&CoreCreateObject{factoryName, typeName, version, &props, newId},
+	)
+}
+
+// CoreDestroy is sent when the client requests to destroy an object.
+type CoreDestroy struct {
+	// The proxy id of the object to destroy.
+	ID Int `json:"id"`
+}
+
+// Opcode satisfies [Message] with a constant value.
+func (c *CoreDestroy) Opcode() byte { return PW_CORE_METHOD_DESTROY }
+
+// FileCount satisfies [Message] with a constant value.
+func (c *CoreDestroy) FileCount() Int { return 0 }
+
+// Size satisfies [KnownSize] with a constant value.
+func (c *CoreDestroy) Size() Word { return SizePrefix + Size(SizeInt) }
+
+// MarshalBinary satisfies [encoding.BinaryMarshaler] via [Marshal].
+func (c *CoreDestroy) MarshalBinary() ([]byte, error) { return Marshal(c) }
+
+// UnmarshalBinary satisfies [encoding.BinaryUnmarshaler] via [Unmarshal].
+func (c *CoreDestroy) UnmarshalBinary(data []byte) error { return Unmarshal(data, c) }
+
+// destroy queues a [CoreDestroy] message for the PipeWire server.
+// This is not safe to use directly, callers should use the exported method
+// on the proxy implementation instead.
+func (core *Core) destroy(id Int) error {
+	return core.ctx.writeMessage(
+		PW_ID_CORE,
+		&CoreDestroy{id},
+	)
+}
+
 // A RegistryGlobal event is emitted to notify a client about a new global object.
 type RegistryGlobal struct {
 	// The global id.
@@ -533,6 +730,30 @@ func (c *RegistryGlobal) MarshalBinary() ([]byte, error) { return Marshal(c) }
 // UnmarshalBinary satisfies [encoding.BinaryUnmarshaler] via [Unmarshal].
 func (c *RegistryGlobal) UnmarshalBinary(data []byte) error { return Unmarshal(data, c) }
 
+// A RegistryGlobalRemove event is emitted when a global with id was removed.
+type RegistryGlobalRemove struct {
+	// The global id that was removed.
+	ID Int `json:"id"`
+}
+
+// Opcode satisfies [Message] with a constant value.
+func (c *RegistryGlobalRemove) Opcode() byte { return PW_REGISTRY_EVENT_GLOBAL_REMOVE }
+
+// FileCount satisfies [Message] with a constant value.
+func (c *RegistryGlobalRemove) FileCount() Int { return 0 }
+
+// Size satisfies [KnownSize] with a constant value.
+func (c *RegistryGlobalRemove) Size() Word {
+	return SizePrefix +
+		Size(SizeInt)
+}
+
+// MarshalBinary satisfies [encoding.BinaryMarshaler] via [Marshal].
+func (c *RegistryGlobalRemove) MarshalBinary() ([]byte, error) { return Marshal(c) }
+
+// UnmarshalBinary satisfies [encoding.BinaryUnmarshaler] via [Unmarshal].
+func (c *RegistryGlobalRemove) UnmarshalBinary(data []byte) error { return Unmarshal(data, c) }
+
 // RegistryBind is sent when the client requests to bind to the
 // global object with id and use the client proxy with new_id as
 // the proxy. After this call, methods can be sent to the remote
@@ -584,10 +805,70 @@ func (registry *Registry) bind(proxy eventProxy, id, version Int) (Int, error) {
 	)
 }
 
+// RegistryDestroy is sent to try to destroy the global object with id.
+// This might fail when the client does not have permission.
+type RegistryDestroy struct {
+	// The global id to destroy.
+	ID Int `json:"id"`
+}
+
+// Opcode satisfies [Message] with a constant value.
+func (c *RegistryDestroy) Opcode() byte { return PW_REGISTRY_METHOD_DESTROY }
+
+// FileCount satisfies [Message] with a constant value.
+func (c *RegistryDestroy) FileCount() Int { return 0 }
+
+// Size satisfies [KnownSize] with a constant value.
+func (c *RegistryDestroy) Size() Word {
+	return SizePrefix +
+		Size(SizeInt)
+}
+
+// MarshalBinary satisfies [encoding.BinaryMarshaler] via [Marshal].
+func (c *RegistryDestroy) MarshalBinary() ([]byte, error) { return Marshal(c) }
+
+// UnmarshalBinary satisfies [encoding.BinaryUnmarshaler] via [Unmarshal].
+func (c *RegistryDestroy) UnmarshalBinary(data []byte) error { return Unmarshal(data, c) }
+
+// destroy queues a [RegistryDestroy] message for the PipeWire server.
+func (registry *Registry) destroy(id Int) error {
+	return registry.ctx.writeMessage(
+		registry.ID,
+		&RegistryDestroy{id},
+	)
+}
+
+// Destroy tries to destroy the global object with id.
+func (registry *Registry) Destroy(id Int) (err error) {
+	asCoreError := registry.ctx.expectsCoreError(registry.ID, &err)
+	if err != nil {
+		return
+	}
+	if err = registry.destroy(id); err != nil {
+		return err
+	}
+	if err = registry.ctx.GetCore().Sync(); err == nil {
+		return nil
+	}
+
+	if coreError := asCoreError(); coreError == nil {
+		return
+	} else {
+		switch syscall.Errno(-coreError.Result) {
+		case syscall.EPERM:
+			return &PermissionError{registry.ID, coreError.Message}
+
+		default:
+			return coreError
+		}
+	}
+}
+
 // An UnsupportedObjectTypeError is the name of a type not known by the server [Registry].
 type UnsupportedObjectTypeError string
 
-func (e UnsupportedObjectTypeError) Error() string { return "unsupported object type " + string(e) }
+func (e UnsupportedObjectTypeError) Error() string   { return "unsupported object type " + string(e) }
+func (e UnsupportedObjectTypeError) Message() string { return e.Error() }
 
 // Core holds state of [PW_TYPE_INTERFACE_Core].
 type Core struct {
@@ -598,22 +879,24 @@ type Core struct {
 	done bool
 
 	ctx *Context
+
 	noAck
+	noRemove
 }
 
 // ErrUnexpectedDone is a [CoreDone] event with unexpected values.
 var ErrUnexpectedDone = errors.New("multiple Core::Done events targeting Core::Sync")
 
-// An UnknownBoundIdError describes the server claiming to have bound a proxy id that was never allocated.
-type UnknownBoundIdError[E any] struct {
+// An UnknownProxyIdError describes an event targeting a proxy id that was never allocated.
+type UnknownProxyIdError[E any] struct {
 	// Offending id decoded from Data.
 	Id Int
 	// Event received from the server.
 	Event E
 }
 
-func (e *UnknownBoundIdError[E]) Error() string {
-	return "unknown bound proxy id " + strconv.Itoa(int(e.Id))
+func (e *UnknownProxyIdError[E]) Error() string {
+	return "unknown proxy id " + strconv.Itoa(int(e.Id))
 }
 
 // An InvalidPingError is a [CorePing] event targeting a proxy id that was never allocated.
@@ -668,6 +951,19 @@ func (core *Core) consume(opcode byte, files []int, unmarshal func(v any)) error
 		unmarshal(&coreError)
 		return &coreError
 
+	case PW_CORE_EVENT_REMOVE_ID:
+		var coreRemoveId CoreRemoveId
+		unmarshal(&coreRemoveId)
+		if proxy, ok := core.ctx.proxy[coreRemoveId.ID]; !ok {
+			// this should never happen so is non-recoverable if it does
+			panic(&UnknownProxyIdError[*CoreRemoveId]{Id: coreRemoveId.ID, Event: &coreRemoveId})
+		} else {
+			delete(core.ctx.proxy, coreRemoveId.ID)
+			// not always populated so this is not checked
+			delete(core.ctx.pendingDestruction, coreRemoveId.ID)
+			return proxy.remove()
+		}
+
 	case PW_CORE_EVENT_BOUND_PROPS:
 		var boundProps CoreBoundProps
 		unmarshal(&boundProps)
@@ -675,12 +971,12 @@ func (core *Core) consume(opcode byte, files []int, unmarshal func(v any)) error
 		delete(core.ctx.pendingIds, boundProps.ID)
 		proxy, ok := core.ctx.proxy[boundProps.ID]
 		if !ok {
-			return &UnknownBoundIdError[*CoreBoundProps]{Id: boundProps.ID, Event: &boundProps}
+			return &UnknownProxyIdError[*CoreBoundProps]{Id: boundProps.ID, Event: &boundProps}
 		}
 		return proxy.setBoundProps(&boundProps)
 
 	default:
-		return &UnsupportedOpcodeError{opcode, core.String()}
+		panic(&UnsupportedOpcodeError{opcode, core.String()})
 	}
 }
 
@@ -698,7 +994,9 @@ type Registry struct {
 	Objects map[Int]RegistryGlobal `json:"objects"`
 
 	ctx *Context
+
 	noAck
+	noRemove
 }
 
 // A GlobalIDCollisionError describes a [RegistryGlobal] event stepping on a previous instance of itself.
@@ -714,6 +1012,14 @@ func (e *GlobalIDCollisionError) Error() string {
 		" stepping on previous id " + strconv.Itoa(int(e.ID)) + " for " + e.Previous.Type
 }
 
+// An UnknownGlobalIDRemoveError describes a [RegistryGlobalRemove] event announcing the removal of
+// a global id that is not yet known to [Registry] or was already deleted.
+type UnknownGlobalIDRemoveError Int
+
+func (e UnknownGlobalIDRemoveError) Error() string {
+	return "Registry::GlobalRemove event targets unknown id " + strconv.Itoa(int(e))
+}
+
 func (registry *Registry) consume(opcode byte, files []int, unmarshal func(v any)) error {
 	closeReceivedFiles(files...)
 	switch opcode {
@@ -727,8 +1033,21 @@ func (registry *Registry) consume(opcode byte, files []int, unmarshal func(v any
 		registry.Objects[global.ID] = global
 		return nil
 
+	case PW_REGISTRY_EVENT_GLOBAL_REMOVE:
+		var globalRemove RegistryGlobalRemove
+		unmarshal(&globalRemove)
+		// server emits PW_CORE_EVENT_REMOVE_ID events targeting
+		// affected proxies so they do not need to be handled here
+		l := len(registry.Objects)
+		delete(registry.Objects, globalRemove.ID)
+		if len(registry.Objects) != l-1 {
+			// this should never happen so is non-recoverable if it does
+			panic(UnknownGlobalIDRemoveError(globalRemove.ID))
+		}
+		return nil
+
 	default:
-		return &UnsupportedOpcodeError{opcode, registry.String()}
+		panic(&UnsupportedOpcodeError{opcode, registry.String()})
 	}
 }
 

internal/pipewire/core_test.go

@@ -171,7 +171,7 @@ func TestCoreError(t *testing.T) {
 			/* padding */ 0, 0, 0, 0,
 
 			/* size: 0x1b bytes */ 0x1b, 0, 0, 0,
-			/*type: String*/ 8, 0, 0, 0,
+			/* type: String */ 8, 0, 0, 0,
 
 			// value: "no permission to destroy 0\x00"
 			0x6e, 0x6f, 0x20, 0x70,
@@ -192,6 +192,24 @@ func TestCoreError(t *testing.T) {
 	}.run(t)
 }
 
+func TestCoreRemoveId(t *testing.T) {
+	t.Parallel()
+
+	encodingTestCases[pipewire.CoreRemoveId, *pipewire.CoreRemoveId]{
+		{"sample", []byte{
+			/* size: rest of data */ 0x10, 0, 0, 0,
+			/* type: Struct */ 0xe, 0, 0, 0,
+
+			/* size: 4 bytes */ 4, 0, 0, 0,
+			/* type: Int */ 4, 0, 0, 0,
+			/* value: 3 */ 3, 0, 0, 0,
+			/* padding */ 0, 0, 0, 0,
+		}, pipewire.CoreRemoveId{
+			ID: 3,
+		}, nil},
+	}.run(t)
+}
+
 func TestCoreBoundProps(t *testing.T) {
 	t.Parallel()
 
@@ -288,6 +306,83 @@ func TestCoreGetRegistry(t *testing.T) {
 	}.run(t)
 }
 
+func TestCoreCreateObject(t *testing.T) {
+	t.Parallel()
+
+	encodingTestCases[pipewire.CoreCreateObject, *pipewire.CoreCreateObject]{
+		{"sample", []byte{
+			/* size: rest of data */ 0x80, 0, 0, 0,
+			/* type: Struct */ 0xe, 0, 0, 0,
+
+			/* size: 0x13 bytes */ 0x13, 0, 0, 0,
+			/* type: String */ 8, 0, 0, 0,
+
+			// value: "spa-device-factory\x00"
+			0x73, 0x70, 0x61, 0x2d,
+			0x64, 0x65, 0x76, 0x69,
+			0x63, 0x65, 0x2d, 0x66,
+			0x61, 0x63, 0x74, 0x6f,
+			0x72, 0x79, 0, 0,
+			0, 0, 0, 0,
+
+			/* size: 0x1a bytes */ 0x1a, 0, 0, 0,
+			/* type: String */ 8, 0, 0, 0,
+
+			// value: "PipeWire:Interface:Device\x00"
+			0x50, 0x69, 0x70, 0x65,
+			0x57, 0x69, 0x72, 0x65,
+			0x3a, 0x49, 0x6e, 0x74,
+			0x65, 0x72, 0x66, 0x61,
+			0x63, 0x65, 0x3a, 0x44,
+			0x65, 0x76, 0x69, 0x63,
+			0x65, 0, 0, 0,
+			0, 0, 0, 0,
+
+			/* size: 4 bytes */ 4, 0, 0, 0,
+			/* type: Int */ 4, 0, 0, 0,
+			/* value: 3 */ 3, 0, 0, 0,
+			/* padding */ 0, 0, 0, 0,
+
+			/* size */ 0x10, 0, 0, 0,
+			/* type: Struct */ 0xe, 0, 0, 0,
+
+			/* size: 4 bytes */ 4, 0, 0, 0,
+			/* type: Int */ 4, 0, 0, 0,
+			/* value: 0 */ 0, 0, 0, 0,
+			/* padding */ 0, 0, 0, 0,
+
+			/* size: 4 bytes */ 4, 0, 0, 0,
+			/* type: Int */ 4, 0, 0, 0,
+			/* value: 0xbad */ 0xad, 0xb, 0, 0,
+			/* padding */ 0, 0, 0, 0,
+		}, pipewire.CoreCreateObject{
+			FactoryName: "spa-device-factory",
+			Type:        pipewire.PW_TYPE_INTERFACE_Device,
+			Version:     pipewire.PW_VERSION_FACTORY,
+			Properties:  &pipewire.SPADict{},
+			NewID:       0xbad,
+		}, nil},
+	}.run(t)
+}
+
+func TestCoreDestroy(t *testing.T) {
+	t.Parallel()
+
+	encodingTestCases[pipewire.CoreDestroy, *pipewire.CoreDestroy]{
+		{"sample", []byte{
+			/* size: rest of data */ 0x10, 0, 0, 0,
+			/* type: Struct */ 0xe, 0, 0, 0,
+
+			/* size: 4 bytes */ 4, 0, 0, 0,
+			/* type: Int */ 4, 0, 0, 0,
+			/* value: 3 */ 3, 0, 0, 0,
+			/* padding */ 0, 0, 0, 0,
+		}, pipewire.CoreDestroy{
+			ID: 3,
+		}, nil},
+	}.run(t)
+}
+
 func TestRegistryGlobal(t *testing.T) {
 	t.Parallel()
 
@@ -745,6 +840,23 @@ func TestRegistryGlobal(t *testing.T) {
 	}.run(t)
 }
 
+func TestRegistryGlobalRemove(t *testing.T) {
+	t.Parallel()
+
+	encodingTestCases[pipewire.RegistryGlobalRemove, *pipewire.RegistryGlobalRemove]{
+		{"sample", []byte{
+			/* size: rest of data*/ 0x10, 0, 0, 0,
+			/* type: Struct */ 0xe, 0, 0, 0,
+			/* size: 4 bytes */ 4, 0, 0, 0,
+			/* type: Int */ 4, 0, 0, 0,
+			/* value: 0xbad */ 0xad, 0xb, 0, 0,
+			/* padding */ 0, 0, 0, 0,
+		}, pipewire.RegistryGlobalRemove{
+			ID: 0xbad,
+		}, nil},
+	}.run(t)
+}
+
 func TestRegistryBind(t *testing.T) {
 	t.Parallel()
 
@@ -757,3 +869,20 @@ func TestRegistryBind(t *testing.T) {
 		}, nil},
 	}.run(t)
 }
+
+func TestRegistryDestroy(t *testing.T) {
+	t.Parallel()
+
+	encodingTestCases[pipewire.RegistryDestroy, *pipewire.RegistryDestroy]{
+		{"sample", []byte{
+			/* size: rest of data*/ 0x10, 0, 0, 0,
+			/* type: Struct */ 0xe, 0, 0, 0,
+			/* size: 4 bytes */ 4, 0, 0, 0,
+			/* type: Int */ 4, 0, 0, 0,
+			/* value: 0xbad */ 0xad, 0xb, 0, 0,
+			/* padding */ 0, 0, 0, 0,
+		}, pipewire.RegistryDestroy{
+			ID: 0xbad,
+		}, nil},
+	}.run(t)
+}

internal/pipewire/pipewire.go

@@ -16,9 +16,9 @@ package pipewire
 
 import (
 	"encoding/binary"
+	"errors"
 	"fmt"
 	"io"
-	"net"
 	"os"
 	"path"
 	"runtime"
@@ -26,10 +26,16 @@ import (
 	"strconv"
 	"strings"
 	"syscall"
+	"time"
 )
 
 // Conn is a low level unix socket interface used by [Context].
 type Conn interface {
+	// MightBlock informs the implementation that the next call to
+	// Recvmsg or Sendmsg might block. A zero or negative timeout
+	// cancels this behaviour.
+	MightBlock(timeout time.Duration)
+
 	// Recvmsg calls syscall.Recvmsg on the underlying socket.
 	Recvmsg(p, oob []byte, flags int) (n, oobn, recvflags int, err error)
 
@@ -51,20 +57,27 @@ type Context struct {
 	buf []byte
 	// Current [Header.Sequence] value, incremented every write.
 	sequence Int
-	// Current server-side [Header.Sequence] value, incremented on every event processed.
-	remoteSequence Int
+	// Pending file descriptors to be sent with the next message.
+	pendingFiles []int
+	// File count already kept track of in [Header].
+	headerFiles int
+
 	// Proxy id associations.
 	proxy map[Int]eventProxy
 	// Newly allocated proxies pending acknowledgement from the server.
 	pendingIds map[Int]struct{}
 	// Smallest available Id for the next proxy.
 	nextId Int
-	// Server side registry generation number.
-	generation Long
-	// Pending file descriptors to be sent with the next message.
-	pendingFiles []int
-	// File count kept track of in [Header].
-	headerFiles int
+	// Proxies targeted by the [CoreDestroy] event pending until next [CoreSync].
+	pendingDestruction map[Int]struct{}
+
+	// Proxy for built-in core events.
+	core Core
+	// Proxy for built-in client events.
+	client Client
+
+	// Current server-side [Header.Sequence] value, incremented on every event processed.
+	remoteSequence Int
 	// Files from the server. This is discarded on every Roundtrip so eventProxy
 	// implementations must make sure to close them to avoid leaking fds.
 	//
@@ -80,13 +93,11 @@ type Context struct {
 	// Pending footer value deferred to the next round trip,
 	// sent if pendingFooter is nil. This is for emulating upstream behaviour
 	deferredPendingFooter KnownSize
+	// Server side registry generation number.
+	generation Long
 	// Deferred operations ran after a [Core.Sync] completes or Close is called. Errors
 	//are reported as part of [ProxyConsumeError] and is not considered fatal unless panicked.
 	syncComplete []func() error
-	// Proxy for built-in core events.
-	core Core
-	// Proxy for built-in client events.
-	client Client
 
 	// Passed to [Conn.Recvmsg]. Not copied if sufficient for all received messages.
 	iovecBuf [1 << 15]byte
@@ -120,8 +131,9 @@ func New(conn Conn, props SPADict) (*Context, error) {
 		PW_ID_CLIENT: {},
 	}
 	ctx.nextId = Int(len(ctx.proxy))
+	ctx.pendingDestruction = make(map[Int]struct{})
 
-	if err := ctx.coreHello(); err != nil {
+	if err := ctx.core.hello(); err != nil {
 		return nil, err
 	}
 	if err := ctx.clientUpdateProperties(props); err != nil {
@@ -131,45 +143,142 @@ func New(conn Conn, props SPADict) (*Context, error) {
 	return &ctx, nil
 }
 
-// A SyscallConnCloser is a [syscall.Conn] that implements [io.Closer].
-type SyscallConnCloser interface {
-	syscall.Conn
-	io.Closer
+// unixConn is an implementation of the [Conn] interface for connections
+// to Unix domain sockets.
+type unixConn struct {
+	fd int
+
+	// Whether creation of a new epoll instance was attempted.
+	epoll bool
+	// File descriptor referring to the new epoll instance.
+	// Valid if epoll is true and epollErr is nil.
+	epollFd int
+	// Error returned by syscall.EpollCreate1.
+	epollErr error
+	// Stores epoll events from the kernel.
+	epollBuf [32]syscall.EpollEvent
+
+	// If non-zero, next call is treated as a blocking call.
+	timeout time.Duration
 }
 
-// A SyscallConn is a [Conn] adapter for [syscall.Conn].
-type SyscallConn struct{ SyscallConnCloser }
+// Dial connects to a Unix domain socket described by name.
+func Dial(name string) (Conn, error) {
+	if fd, err := syscall.Socket(syscall.AF_UNIX, syscall.SOCK_STREAM|syscall.SOCK_CLOEXEC|syscall.SOCK_NONBLOCK, 0); err != nil {
+		return nil, os.NewSyscallError("socket", err)
+	} else if err = syscall.Connect(fd, &syscall.SockaddrUnix{Name: name}); err != nil {
+		_ = syscall.Close(fd)
+		return nil, os.NewSyscallError("connect", err)
+	} else {
+		return &unixConn{fd: fd}, nil
+	}
+}
 
-// Recvmsg implements [Conn.Recvmsg] via [syscall.Conn.SyscallConn].
-func (conn SyscallConn) Recvmsg(p, oob []byte, flags int) (n, oobn, recvflags int, err error) {
-	var rc syscall.RawConn
-	if rc, err = conn.SyscallConn(); err != nil {
+// MightBlock informs the implementation that the next call
+// might block for a non-zero timeout.
+func (conn *unixConn) MightBlock(timeout time.Duration) {
+	if timeout < 0 {
+		timeout = 0
+	}
+	conn.timeout = timeout
+}
+
+// wantsEpoll is called at the beginning of any method that might use epoll.
+func (conn *unixConn) wantsEpoll() error {
+	if !conn.epoll {
+		conn.epoll = true
+		conn.epollFd, conn.epollErr = syscall.EpollCreate1(syscall.EPOLL_CLOEXEC)
+		if conn.epollErr == nil {
+			if conn.epollErr = syscall.EpollCtl(conn.epollFd, syscall.EPOLL_CTL_ADD, conn.fd, &syscall.EpollEvent{
+				Events: syscall.EPOLLERR | syscall.EPOLLHUP,
+				Fd:     int32(conn.fd),
+			}); conn.epollErr != nil {
+				_ = syscall.Close(conn.epollFd)
+			}
+		}
+	}
+	return conn.epollErr
+}
+
+// wait waits for a specific I/O event on fd. Caller must arrange for wantsEpoll
+// to be called somewhere before wait is called.
+func (conn *unixConn) wait(event uint32) (err error) {
+	if conn.timeout == 0 {
+		return nil
+	}
+	deadline := time.Now().Add(conn.timeout)
+	conn.timeout = 0
+
+	if err = syscall.EpollCtl(conn.epollFd, syscall.EPOLL_CTL_MOD, conn.fd, &syscall.EpollEvent{
+		Events: event | syscall.EPOLLERR | syscall.EPOLLHUP,
+		Fd:     int32(conn.fd),
+	}); err != nil {
 		return
 	}
 
-	if controlErr := rc.Control(func(fd uintptr) {
-		n, oobn, recvflags, _, err = syscall.Recvmsg(int(fd), p, oob, flags)
-	}); controlErr != nil && err == nil {
-		err = controlErr
+	for timeout := deadline.Sub(time.Now()); timeout > 0; timeout = deadline.Sub(time.Now()) {
+		var n int
+		if n, err = syscall.EpollWait(conn.epollFd, conn.epollBuf[:], int(timeout/time.Millisecond)); err != nil {
+			return
+		}
+
+		switch n {
+		case 1: // only the socket fd is ever added
+			if conn.epollBuf[0].Fd != int32(conn.fd) { // unreachable
+				return syscall.ENOTRECOVERABLE
+			}
+			if conn.epollBuf[0].Events&event == event ||
+				conn.epollBuf[0].Events&syscall.EPOLLERR|syscall.EPOLLHUP != 0 {
+				return nil
+			}
+			err = syscall.ETIME
+			continue
+
+		case 0: // timeout
+			return syscall.ETIMEDOUT
+
+		default: // unreachable
+			return syscall.ENOTRECOVERABLE
+		}
 	}
 	return
 }
 
-// Sendmsg implements [Conn.Sendmsg] via [syscall.Conn.SyscallConn].
-func (conn SyscallConn) Sendmsg(p, oob []byte, flags int) (n int, err error) {
-	var rc syscall.RawConn
-	if rc, err = conn.SyscallConn(); err != nil {
+// Recvmsg calls syscall.Recvmsg on the underlying socket.
+func (conn *unixConn) Recvmsg(p, oob []byte, flags int) (n, oobn, recvflags int, err error) {
+	if err = conn.wantsEpoll(); err != nil {
+		return
+	} else if err = conn.wait(syscall.EPOLLIN); err != nil {
 		return
 	}
 
-	if controlErr := rc.Control(func(fd uintptr) {
-		n, err = syscall.SendmsgN(int(fd), p, oob, nil, flags)
-	}); controlErr != nil && err == nil {
-		err = controlErr
-	}
+	n, oobn, recvflags, _, err = syscall.Recvmsg(conn.fd, p, oob, flags)
 	return
 }
 
+// Sendmsg calls syscall.Sendmsg on the underlying socket.
+func (conn *unixConn) Sendmsg(p, oob []byte, flags int) (n int, err error) {
+	if err = conn.wantsEpoll(); err != nil {
+		return
+	} else if err = conn.wait(syscall.EPOLLOUT); err != nil {
+		return
+	}
+
+	n, err = syscall.SendmsgN(conn.fd, p, oob, nil, flags)
+	return
+}
+
+// Close closes the underlying socket and the epoll fd if populated.
+func (conn *unixConn) Close() (err error) {
+	if conn.epoll && conn.epollErr == nil {
+		conn.epollErr = syscall.Close(conn.epollFd)
+	}
+	if err = syscall.Close(conn.fd); err != nil {
+		return
+	}
+	return conn.epollErr
+}
+
 // MustNew calls [New](conn, props) and panics on error.
 // It is intended for use in tests with hard-coded strings.
 func MustNew(conn Conn, props SPADict) *Context {
@@ -271,16 +380,20 @@ func (ctx *Context) recvmsg(remaining []byte) (payload []byte, err error) {
 		n, oobn, recvflags, err = ctx.conn.Recvmsg(ctx.iovecBuf[len(remaining):], ctx.oobBuf[:], recvmsgFlags)
 
 		if oob := ctx.oobBuf[:oobn]; len(oob) > 0 {
+			var oobErr error
+
 			var scm []syscall.SocketControlMessage
-			if scm, err = syscall.ParseSocketControlMessage(oob); err != nil {
+			if scm, oobErr = syscall.ParseSocketControlMessage(oob); oobErr != nil {
 				ctx.closeReceivedFiles()
+				err = oobErr
 				return
 			}
 
 			var fds []int
 			for i := range scm {
-				if fds, err = syscall.ParseUnixRights(&scm[i]); err != nil {
+				if fds, oobErr = syscall.ParseUnixRights(&scm[i]); oobErr != nil {
 					ctx.closeReceivedFiles()
+					err = oobErr
 					return
 				}
 				ctx.receivedFiles = append(ctx.receivedFiles, fds...)
@@ -299,7 +412,7 @@ func (ctx *Context) recvmsg(remaining []byte) (payload []byte, err error) {
 			}
 			if err != syscall.EAGAIN && err != syscall.EWOULDBLOCK {
 				ctx.closeReceivedFiles()
-				return nil, os.NewSyscallError("recvmsg", err)
+				return nil, &ProxyFatalError{Err: os.NewSyscallError("recvmsg", err), ProxyErrs: ctx.cloneAsProxyErrors()}
 			}
 		}
 
@@ -310,7 +423,7 @@ func (ctx *Context) recvmsg(remaining []byte) (payload []byte, err error) {
 		err = syscall.EPIPE // not wrapped as it did not come from the syscall
 	}
 	if n > 0 {
-		payload = ctx.iovecBuf[:n]
+		payload = ctx.iovecBuf[:len(remaining)+n]
 	}
 	return
 }
@@ -336,7 +449,7 @@ func (ctx *Context) sendmsg(p []byte, fds ...int) error {
 		}
 
 		if err != nil && err != syscall.EAGAIN && err != syscall.EWOULDBLOCK {
-			return os.NewSyscallError("sendmsg", err)
+			return &ProxyFatalError{Err: os.NewSyscallError("sendmsg", err), ProxyErrs: ctx.cloneAsProxyErrors()}
 		}
 		return err
 	}
@@ -415,6 +528,9 @@ type eventProxy interface {
 	consume(opcode byte, files []int, unmarshal func(v any)) error
 	// setBoundProps stores a [CoreBoundProps] event received from the server.
 	setBoundProps(event *CoreBoundProps) error
+	// remove is called when the proxy is removed for any reason, usually from
+	// being targeted by a [PW_CORE_EVENT_REMOVE_ID] event.
+	remove() error
 
 	// Stringer returns the PipeWire interface name.
 	fmt.Stringer
@@ -495,13 +611,21 @@ func (e DanglingFilesError) Error() string {
 }
 
 // An UnacknowledgedProxyError holds newly allocated proxy ids that the server failed
-// to acknowledge after an otherwise successful [Context.Roundtrip].
+// to acknowledge after an otherwise successful [Core.Sync].
 type UnacknowledgedProxyError []Int
 
 func (e UnacknowledgedProxyError) Error() string {
 	return "server did not acknowledge " + strconv.Itoa(len(e)) + " proxies"
 }
 
+// An UnacknowledgedProxyDestructionError holds destroyed proxy ids that the server failed
+// to acknowledge after an otherwise successful [Core.Sync].
+type UnacknowledgedProxyDestructionError []Int
+
+func (e UnacknowledgedProxyDestructionError) Error() string {
+	return "server did not acknowledge " + strconv.Itoa(len(e)) + " proxy destructions"
+}
+
 // A ProxyFatalError describes an error that terminates event handling during a
 // [Context.Roundtrip] and makes further event processing no longer possible.
 type ProxyFatalError struct {
@@ -576,11 +700,21 @@ func (ctx *Context) Roundtrip() (err error) {
 	return
 }
 
+const (
+	// roundtripTimeout is the maximum duration socket operations during
+	// Context.roundtrip is allowed to block for.
+	roundtripTimeout = 5 * time.Second
+)
+
 // roundtrip implements the Roundtrip method without checking proxyErrors.
 func (ctx *Context) roundtrip() (err error) {
+	ctx.conn.MightBlock(roundtripTimeout)
 	if err = ctx.sendmsg(ctx.buf, ctx.pendingFiles...); err != nil {
 		return
 	}
+	ctx.buf = ctx.buf[:0]
+	ctx.pendingFiles = ctx.pendingFiles[:0]
+	ctx.headerFiles = 0
 
 	defer func() {
 		var danglingFiles DanglingFilesError
@@ -608,6 +742,7 @@ func (ctx *Context) roundtrip() (err error) {
 	}()
 
 	var remaining []byte
+	ctx.conn.MightBlock(roundtripTimeout)
 	for {
 		remaining, err = ctx.consume(remaining)
 		if err == nil {
@@ -629,7 +764,7 @@ func (ctx *Context) roundtrip() (err error) {
 }
 
 // currentSeq returns the current sequence number.
-// This must only be called from eventProxy.consume.
+// This must only be called immediately after queueing a message.
 func (ctx *Context) currentSeq() Int { return ctx.sequence - 1 }
 
 // currentRemoteSeq returns the current remote sequence number.
@@ -657,10 +792,6 @@ func (ctx *Context) consume(receiveRemaining []byte) (remaining []byte, err erro
 		return
 	}()
 
-	ctx.buf = ctx.buf[:0]
-	ctx.pendingFiles = ctx.pendingFiles[:0]
-	ctx.headerFiles = 0
-
 	if remaining, err = ctx.recvmsg(receiveRemaining); err != nil {
 		return
 	}
@@ -677,11 +808,11 @@ func (ctx *Context) consume(receiveRemaining []byte) (remaining []byte, err erro
 		if header.Sequence != ctx.remoteSequence {
 			return remaining, UnexpectedSequenceError(header.Sequence)
 		}
-		ctx.remoteSequence++
 
 		if len(remaining) < int(SizeHeader+header.Size) {
 			return
 		}
+		ctx.remoteSequence++
 
 		proxy, ok := ctx.proxy[header.ID]
 		if !ok {
@@ -783,18 +914,67 @@ func (ctx *Context) Close() (err error) {
 	}
 }
 
+// expectsCoreError returns a function that inspects an error value and
+// returns the address of a [CoreError] if it is the only error present
+// and targets the specified proxy and sequence.
+//
+// The behaviour of expectsCoreError is only correct for an empty buf
+// prior to calling. If buf is not empty, [Core.Sync] is called, with
+// its return value stored to the value pointed to by errP if not nil,
+// and the function is not populated.
+//
+// The caller must queue a message and call [Core.Sync] immediately
+// after calling expectsCoreError.
+func (ctx *Context) expectsCoreError(id Int, errP *error) (asCoreError func() (coreError *CoreError)) {
+	if len(ctx.buf) > 0 {
+		if err := ctx.GetCore().Sync(); err != nil {
+			*errP = err
+			return nil
+		}
+	}
+
+	sequence := ctx.sequence
+	return func() (coreError *CoreError) {
+		if proxyErrors, ok := (*errP).(ProxyConsumeError); !ok ||
+			len(proxyErrors) != 1 ||
+			!errors.As(proxyErrors[0], &coreError) ||
+			coreError == nil ||
+			coreError.ID != id ||
+			coreError.Sequence != sequence {
+			// do not return a non-matching CoreError
+			coreError = nil
+		}
+		return
+	}
+}
+
+// A PermissionError describes an error emitted by the server when trying to
+// perform an operation that the client has no permission for.
+type PermissionError struct {
+	// The id of the resource (proxy if emitted by the client) that is in error.
+	ID Int `json:"id"`
+	// An error message.
+	Message string `json:"message"`
+}
+
+func (*PermissionError) Unwrap() error   { return syscall.EPERM }
+func (e *PermissionError) Error() string { return e.Message }
+
+// Remote is the environment (sic) with the remote name.
+const Remote = "PIPEWIRE_REMOTE"
+
 /* modules/module-protocol-native/local-socket.c */
 
 const DEFAULT_SYSTEM_RUNTIME_DIR = "/run/pipewire"
 
-// connectName connects to a PipeWire remote by name and returns the [net.UnixConn].
-func connectName(name string, manager bool) (conn *net.UnixConn, err error) {
+// connectName connects to a PipeWire remote by name and returns the resulting [Conn].
+func connectName(name string, manager bool) (conn Conn, err error) {
 	if manager && !strings.HasSuffix(name, "-manager") {
 		return connectName(name+"-manager", false)
 	}
 
 	if path.IsAbs(name) || (len(name) > 0 && name[0] == '@') {
-		return net.DialUnix("unix", nil, &net.UnixAddr{Name: name, Net: "unix"})
+		return Dial(name)
 	} else {
 		runtimeDir, ok := os.LookupEnv("PIPEWIRE_RUNTIME_DIR")
 		if !ok || !path.IsAbs(runtimeDir) {
@@ -809,31 +989,29 @@ func connectName(name string, manager bool) (conn *net.UnixConn, err error) {
 		if !ok || !path.IsAbs(runtimeDir) {
 			runtimeDir = DEFAULT_SYSTEM_RUNTIME_DIR
 		}
-		return net.DialUnix("unix", nil, &net.UnixAddr{Name: path.Join(runtimeDir, name), Net: "unix"})
+		return Dial(path.Join(runtimeDir, name))
 	}
 }
 
 // ConnectName connects to a PipeWire remote by name.
-func ConnectName(name string, manager bool) (ctx *Context, err error) {
-	var props SPADict
+func ConnectName(name string, manager bool, props SPADict) (ctx *Context, err error) {
 	if manager {
 		props = append(props, SPADictItem{Key: PW_KEY_REMOTE_INTENTION, Value: "manager"})
 	}
 
 	if name == "" {
-		if v, ok := os.LookupEnv("PIPEWIRE_REMOTE"); !ok || v == "" {
+		if v, ok := os.LookupEnv(Remote); !ok || v == "" {
 			name = PW_DEFAULT_REMOTE
 		} else {
 			name = v
 		}
 	}
 
-	var conn *net.UnixConn
+	var conn Conn
 	if conn, err = connectName(name, manager); err != nil {
 		return
 	}
-
-	if ctx, err = New(SyscallConn{conn}, props); err != nil {
+	if ctx, err = New(conn, props); err != nil {
 		ctx = nil
 		_ = conn.Close()
 	}
@@ -841,4 +1019,6 @@ func ConnectName(name string, manager bool) (ctx *Context, err error) {
 }
 
 // Connect connects to the PipeWire remote.
-func Connect(manager bool) (ctx *Context, err error) { return ConnectName("", manager) }
+func Connect(manager bool, props SPADict) (ctx *Context, err error) {
+	return ConnectName("", manager, props)
+}

internal/pipewire/pipewire_test.go

@@ -6,6 +6,7 @@ import (
 	"strconv"
 	. "syscall"
 	"testing"
+	"time"
 
 	"hakurei.app/container/stub"
 	"hakurei.app/internal/pipewire"
@@ -680,9 +681,6 @@ func TestContext(t *testing.T) {
 	}); err != nil {
 		t.Fatalf("SecurityContext.Create: error = %v", err)
 	}
-	if err := ctx.GetCore().Sync(); err != nil {
-		t.Fatalf("Sync: error = %v", err)
-	}
 
 	// none of these should change
 	if coreInfo := ctx.GetCore().Info; !reflect.DeepEqual(coreInfo, &wantCoreInfo0) {
@@ -718,6 +716,18 @@ type stubUnixConn struct {
 	current int
 }
 
+func (conn *stubUnixConn) MightBlock(timeout time.Duration) {
+	if timeout != 5*time.Second {
+		panic("unexpected timeout " + timeout.String())
+	}
+	if conn.current == 0 ||
+		(conn.samples[conn.current-1].nr == SYS_RECVMSG && conn.samples[conn.current-1].errno == EAGAIN && conn.samples[conn.current].nr == SYS_SENDMSG) ||
+		(conn.samples[conn.current-1].nr == SYS_SENDMSG && conn.samples[conn.current].nr == SYS_RECVMSG) {
+		return
+	}
+	panic("unexpected blocking hint before sample " + strconv.Itoa(conn.current))
+}
+
 // nextSample returns the current sample and increments the counter.
 func (conn *stubUnixConn) nextSample(nr uintptr) (sample *stubUnixConnSample, wantOOB []byte, err error) {
 	sample = &conn.samples[conn.current]
@@ -836,6 +846,7 @@ func TestContextErrors(t *testing.T) {
 
 		{"UnexpectedFileCountError", &pipewire.UnexpectedFileCountError{0, -1}, "received -1 files instead of the expected 0"},
 		{"UnacknowledgedProxyError", make(pipewire.UnacknowledgedProxyError, 1<<4), "server did not acknowledge 16 proxies"},
+		{"UnacknowledgedProxyDestructionError", make(pipewire.UnacknowledgedProxyDestructionError, 1<<4), "server did not acknowledge 16 proxy destructions"},
 		{"DanglingFilesError", make(pipewire.DanglingFilesError, 1<<4), "received 16 dangling files"},
 		{"UnexpectedFilesError", pipewire.UnexpectedFilesError(1 << 4), "server message headers claim to have sent more files than actually received"},
 		{"UnexpectedSequenceError", pipewire.UnexpectedSequenceError(1 << 4), "unexpected seq 16"},
@@ -862,6 +873,19 @@ func TestContextErrors(t *testing.T) {
 			ID:       0xbad,
 			Sequence: 0xcafe,
 		}, "received Core::Ping seq 51966 targeting unknown proxy id 2989"},
+
+		{"GlobalIDCollisionError", &pipewire.GlobalIDCollisionError{
+			ID:       0xbad,
+			Previous: &pipewire.RegistryGlobal{Type: "PipeWire:Interface:Invalid"},
+			Current:  &pipewire.RegistryGlobal{Type: "PipeWire:Interface:NewInvalid"},
+		}, "new Registry::Global event for PipeWire:Interface:NewInvalid stepping on previous id 2989 for PipeWire:Interface:Invalid"},
+
+		{"UnknownGlobalIDRemoveError", pipewire.UnknownGlobalIDRemoveError(0xbad), "Registry::GlobalRemove event targets unknown id 2989"},
+
+		{"PermissionError", &pipewire.PermissionError{
+			ID:      2,
+			Message: "no permission to destroy 0",
+		}, "no permission to destroy 0"},
 	}
 	for _, tc := range testCases {
 		t.Run(tc.name, func(t *testing.T) {

internal/pipewire/pod.go

@@ -384,19 +384,16 @@ func unmarshalValue(data []byte, v reflect.Value, wireSizeP *Word) error {
 		return nil
 
 	case reflect.Pointer:
-		if len(data) < SizePrefix {
-			return ErrEOFPrefix
-		}
-		switch SPAKind(binary.NativeEndian.Uint32(data[SizeSPrefix:])) {
-		case SPA_TYPE_None:
+		if ok, err := unmarshalHandleNone(&data, wireSizeP); err != nil {
+			return err
+		} else if ok {
 			v.SetZero()
 			return nil
-
-		default:
-			v.Set(reflect.New(v.Type().Elem()))
-			return unmarshalValue(data, v.Elem(), wireSizeP)
 		}
 
+		v.Set(reflect.New(v.Type().Elem()))
+		return unmarshalValue(data, v.Elem(), wireSizeP)
+
 	case reflect.String:
 		*wireSizeP = 0
 		if err := unmarshalCheckTypeBounds(&data, SPA_TYPE_String, wireSizeP); err != nil {
@@ -422,6 +419,29 @@ func unmarshalValue(data []byte, v reflect.Value, wireSizeP *Word) error {
 	}
 }
 
+// unmarshalHandleNone establishes prefix bounds and, for a value of type [SPA_TYPE_None],
+// validates its size and skips the header. This is for unmarshalling values that can be nil.
+func unmarshalHandleNone(data *[]byte, wireSizeP *Word) (bool, error) {
+	if len(*data) < SizePrefix {
+		return false, ErrEOFPrefix
+	}
+
+	if SPAKind(binary.NativeEndian.Uint32((*data)[SizeSPrefix:])) != SPA_TYPE_None {
+		return false, nil
+	}
+
+	*wireSizeP = 0
+	if err := unmarshalCheckTypeBounds(data, SPA_TYPE_None, wireSizeP); err != nil {
+		return true, err
+	}
+
+	if len(*data) != 0 {
+		return true, TrailingGarbageError(*data)
+	}
+
+	return true, nil
+}
+
 // An InconsistentSizeError describes an inconsistent size prefix encountered
 // in data passed to [Unmarshal].
 type InconsistentSizeError struct{ Prefix, Expect Word }

internal/pipewire/securitycontext.go

@@ -26,9 +26,10 @@ const (
 
 const (
 	PW_SECURITY_CONTEXT_METHOD_ADD_LISTENER = iota
-	PW_SECURITY_CONTEXT_METHOD_CREATE
-	PW_SECURITY_CONTEXT_METHOD_NUM
 
+	PW_SECURITY_CONTEXT_METHOD_CREATE
+
+	PW_SECURITY_CONTEXT_METHOD_NUM
 	PW_VERSION_SECURITY_CONTEXT_METHODS = 0
 )
 
@@ -90,6 +91,8 @@ type SecurityContext struct {
 	GlobalID Int `json:"id"`
 
 	ctx *Context
+
+	destructible
 }
 
 // GetSecurityContext queues a [RegistryBind] message for the PipeWire server
@@ -108,13 +111,39 @@ func (registry *Registry) GetSecurityContext() (securityContext *SecurityContext
 }
 
 // Create queues a [SecurityContextCreate] message for the PipeWire server.
-func (securityContext *SecurityContext) Create(listenFd, closeFd int, props SPADict) error {
+func (securityContext *SecurityContext) Create(listenFd, closeFd int, props SPADict) (err error) {
+	if err = securityContext.checkDestroy(); err != nil {
+		return
+	}
+
+	asCoreError := securityContext.ctx.expectsCoreError(securityContext.ID, &err)
+	if err != nil {
+		return
+	}
+
 	// queued in reverse based on upstream behaviour, unsure why
 	offset := securityContext.ctx.queueFiles(closeFd, listenFd)
-	return securityContext.ctx.writeMessage(
+	if err = securityContext.ctx.writeMessage(
 		securityContext.ID,
 		&SecurityContextCreate{ListenFd: offset + 1, CloseFd: offset + 0, Properties: &props},
-	)
+	); err != nil {
+		return
+	}
+	if err = securityContext.ctx.GetCore().Sync(); err == nil {
+		return nil
+	}
+
+	if coreError := asCoreError(); coreError == nil {
+		return
+	} else {
+		switch syscall.Errno(-coreError.Result) {
+		case syscall.EPERM:
+			return &PermissionError{securityContext.ID, coreError.Message}
+
+		default:
+			return coreError
+		}
+	}
 }
 
 // securityContextCloser holds onto resources associated to the security context.
@@ -144,6 +173,9 @@ func (scc *securityContextCloser) Close() (err error) {
 // BindAndCreate binds a new socket to the specified pathname and pass it to Create.
 // It returns an [io.Closer] corresponding to [SecurityContextCreate.CloseFd].
 func (securityContext *SecurityContext) BindAndCreate(pathname string, props SPADict) (io.Closer, error) {
+	if err := securityContext.checkDestroy(); err != nil {
+		return nil, err
+	}
 	var scc securityContextCloser
 
 	// ensure pathname is available
@@ -185,17 +217,19 @@ func (securityContext *SecurityContext) BindAndCreate(pathname string, props SPA
 }
 
 func (securityContext *SecurityContext) consume(opcode byte, files []int, _ func(v any)) error {
+	securityContext.mustCheckDestroy()
 	closeReceivedFiles(files...)
 	switch opcode {
 	// SecurityContext does not receive any events
 
 	default:
-		return &UnsupportedOpcodeError{opcode, securityContext.String()}
+		panic(&UnsupportedOpcodeError{opcode, securityContext.String()})
 	}
 
 }
 
 func (securityContext *SecurityContext) setBoundProps(event *CoreBoundProps) error {
+	securityContext.mustCheckDestroy()
 	if securityContext.ID != event.ID {
 		return &InconsistentIdError{Proxy: securityContext, ID: securityContext.ID, ServerID: event.ID}
 	}
@@ -205,4 +239,9 @@ func (securityContext *SecurityContext) setBoundProps(event *CoreBoundProps) err
 	return nil
 }
 
+// Destroy destroys this [SecurityContext] proxy.
+func (securityContext *SecurityContext) Destroy() error {
+	return securityContext.destroy(securityContext.ctx, securityContext.ID)
+}
+
 func (securityContext *SecurityContext) String() string { return PW_TYPE_INTERFACE_SecurityContext }

internal/pkg/dir.go

@@ -0,0 +1,211 @@
+package pkg
+
+import (
+	"crypto/sha512"
+	"encoding/binary"
+	"errors"
+	"io"
+	"io/fs"
+	"math"
+	"os"
+	"path/filepath"
+	"syscall"
+
+	"hakurei.app/container/check"
+)
+
+// FlatEntry is a directory entry to be encoded for [Flatten].
+type FlatEntry struct {
+	Mode fs.FileMode // file mode bits
+	Path string      // pathname of the file
+	Data []byte      // file content or symlink destination
+}
+
+/*
+| mode uint32 | path_sz uint32 |
+|        data_sz uint64        |
+|         path string          |
+|         data []byte          |
+*/
+
+// wordSize is the boundary which binary segments are always aligned to.
+const wordSize = 8
+
+// alignSize returns the padded size for aligning sz.
+func alignSize(sz int) int {
+	return sz + (wordSize-(sz)%wordSize)%wordSize
+}
+
+// Encode encodes the entry for transmission or hashing.
+func (ent *FlatEntry) Encode(w io.Writer) (n int, err error) {
+	pPathSize := alignSize(len(ent.Path))
+	if pPathSize > math.MaxUint32 {
+		return 0, syscall.E2BIG
+	}
+	pDataSize := alignSize(len(ent.Data))
+
+	payload := make([]byte, wordSize*2+pPathSize+pDataSize)
+	binary.LittleEndian.PutUint32(payload, uint32(ent.Mode))
+	binary.LittleEndian.PutUint32(payload[wordSize/2:], uint32(len(ent.Path)))
+	binary.LittleEndian.PutUint64(payload[wordSize:], uint64(len(ent.Data)))
+	copy(payload[wordSize*2:], ent.Path)
+	copy(payload[wordSize*2+pPathSize:], ent.Data)
+	return w.Write(payload)
+}
+
+// ErrInsecurePath is returned by [FlatEntry.Decode] if validation is requested
+// and a nonlocal path is encountered in the stream.
+var ErrInsecurePath = errors.New("insecure file path")
+
+// Decode decodes the entry from its representation produced by Encode.
+func (ent *FlatEntry) Decode(r io.Reader, validate bool) (n int, err error) {
+	var nr int
+
+	header := make([]byte, wordSize*2)
+	nr, err = r.Read(header)
+	n += nr
+	if err != nil {
+		if errors.Is(err, io.EOF) && n != 0 {
+			err = io.ErrUnexpectedEOF
+		}
+		return
+	}
+
+	ent.Mode = fs.FileMode(binary.LittleEndian.Uint32(header))
+	pathSize := int(binary.LittleEndian.Uint32(header[wordSize/2:]))
+	pPathSize := alignSize(pathSize)
+	dataSize := int(binary.LittleEndian.Uint64(header[wordSize:]))
+	pDataSize := alignSize(dataSize)
+
+	buf := make([]byte, pPathSize+pDataSize)
+	nr, err = r.Read(buf)
+	n += nr
+	if err != nil {
+		if errors.Is(err, io.EOF) {
+			if nr != len(buf) {
+				err = io.ErrUnexpectedEOF
+				return
+			}
+		} else {
+			return
+		}
+	}
+
+	ent.Path = string(buf[:pathSize])
+	if ent.Mode.IsDir() {
+		ent.Data = nil
+	} else {
+		ent.Data = buf[pPathSize : pPathSize+dataSize]
+	}
+
+	if validate && !filepath.IsLocal(ent.Path) {
+		err = ErrInsecurePath
+	}
+
+	return
+}
+
+// DirScanner provides an efficient interface for reading a stream of encoded
+// [FlatEntry]. Successive calls to the Scan method will step through the
+// entries in the stream.
+type DirScanner struct {
+	// Underlying reader to scan [FlatEntry] representations from.
+	r io.Reader
+
+	// First non-EOF I/O error, returned by the Err method.
+	err error
+
+	// Entry to store results in. Its address is returned by the Entry method
+	// and is updated on every call to Scan.
+	ent FlatEntry
+
+	// Validate pathnames during decoding.
+	validate bool
+}
+
+// NewDirScanner returns the address of a new instance of [DirScanner] reading
+// from r. The caller must no longer read from r after this function returns.
+func NewDirScanner(r io.Reader, validate bool) *DirScanner {
+	return &DirScanner{r: r, validate: validate}
+}
+
+// Err returns the first non-EOF I/O error.
+func (s *DirScanner) Err() error {
+	if errors.Is(s.err, io.EOF) {
+		return nil
+	}
+	return s.err
+}
+
+// Entry returns the address to the [FlatEntry] value storing the last result.
+func (s *DirScanner) Entry() *FlatEntry { return &s.ent }
+
+// Scan advances to the next [FlatEntry].
+func (s *DirScanner) Scan() bool {
+	if s.err != nil {
+		return false
+	}
+
+	var n int
+	n, s.err = s.ent.Decode(s.r, s.validate)
+	if errors.Is(s.err, io.EOF) {
+		return n != 0
+	}
+	return s.err == nil
+}
+
+// Flatten writes a deterministic representation of the contents of fsys to w.
+// The resulting data can be hashed to produce a deterministic checksum for the
+// directory.
+func Flatten(fsys fs.FS, root string, w io.Writer) (n int, err error) {
+	var nr int
+	err = fs.WalkDir(fsys, root, func(path string, d fs.DirEntry, err error) error {
+		if err != nil {
+			return err
+		}
+
+		var fi fs.FileInfo
+		fi, err = d.Info()
+		if err != nil {
+			return err
+		}
+
+		ent := FlatEntry{
+			Path: path,
+			Mode: fi.Mode(),
+		}
+		if ent.Mode.IsRegular() {
+			if ent.Data, err = fs.ReadFile(fsys, path); err != nil {
+				return err
+			}
+		} else if ent.Mode&fs.ModeSymlink != 0 {
+			var newpath string
+			if newpath, err = fs.ReadLink(fsys, path); err != nil {
+				return err
+			}
+			ent.Data = []byte(newpath)
+		} else if !ent.Mode.IsDir() {
+			return InvalidFileModeError(ent.Mode)
+		}
+
+		nr, err = ent.Encode(w)
+		n += nr
+		return err
+	})
+	return
+}
+
+// HashFS returns a checksum produced by hashing the result of [Flatten].
+func HashFS(buf *Checksum, fsys fs.FS, root string) error {
+	h := sha512.New384()
+	if _, err := Flatten(fsys, root, h); err != nil {
+		return err
+	}
+	h.Sum(buf[:0])
+	return nil
+}
+
+// HashDir returns a checksum produced by hashing the result of [Flatten].
+func HashDir(buf *Checksum, pathname *check.Absolute) error {
+	return HashFS(buf, os.DirFS(pathname.String()), ".")
+}

internal/pkg/dir_test.go

@@ -0,0 +1,570 @@
+package pkg_test
+
+import (
+	"bytes"
+	"io/fs"
+	"reflect"
+	"testing"
+	"testing/fstest"
+
+	"hakurei.app/internal/pkg"
+)
+
+func TestFlatten(t *testing.T) {
+	t.Parallel()
+
+	testCases := []struct {
+		name    string
+		fsys    fs.FS
+		entries []pkg.FlatEntry
+		sum     pkg.Checksum
+		err     error
+	}{
+		{"bad type", fstest.MapFS{
+			".":       {Mode: fs.ModeDir | 0700},
+			"invalid": {Mode: fs.ModeCharDevice | 0400},
+		}, nil, pkg.Checksum{}, pkg.InvalidFileModeError(
+			fs.ModeCharDevice | 0400,
+		)},
+
+		{"empty", fstest.MapFS{
+			".":          {Mode: fs.ModeDir | 0700},
+			"checksum":   {Mode: fs.ModeDir | 0700},
+			"identifier": {Mode: fs.ModeDir | 0700},
+			"work":       {Mode: fs.ModeDir | 0700},
+		}, []pkg.FlatEntry{
+			{Mode: fs.ModeDir | 0700, Path: "."},
+			{Mode: fs.ModeDir | 0700, Path: "checksum"},
+			{Mode: fs.ModeDir | 0700, Path: "identifier"},
+			{Mode: fs.ModeDir | 0700, Path: "work"},
+		}, pkg.MustDecode("E4vEZKhCcL2gPZ2Tt59FS3lDng-d_2SKa2i5G_RbDfwGn6EemptFaGLPUDiOa94C"), nil},
+
+		{"sample cache file", fstest.MapFS{
+			".": {Mode: fs.ModeDir | 0700},
+
+			"checksum": {Mode: fs.ModeDir | 0700},
+			"checksum/vsAhtPNo4waRNOASwrQwcIPTqb3SBuJOXw2G4T1mNmVZM-wrQTRllmgXqcIIoRcX": {Mode: 0400, Data: []byte{0}},
+			"checksum/0bSFPu5Tnd-2Jj0Mv6co23PW2t3BmHc7eLFj9TgY3eIBg8zislo7xZYNBqovVLcq": {Mode: 0400, Data: []byte{0, 0, 0, 0, 0xad, 0xb, 0, 4, 0xfe, 0xfe, 0, 0, 0xfe, 0xca, 0, 0}},
+
+			"identifier": {Mode: fs.ModeDir | 0700},
+			"identifier/vsAhtPNo4waRNOASwrQwcIPTqb3SBuJOXw2G4T1mNmVZM-wrQTRllmgXqcIIoRcX": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/vsAhtPNo4waRNOASwrQwcIPTqb3SBuJOXw2G4T1mNmVZM-wrQTRllmgXqcIIoRcX")},
+			"identifier/0bSFPu5Tnd-2Jj0Mv6co23PW2t3BmHc7eLFj9TgY3eIBg8zislo7xZYNBqovVLcq": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/0bSFPu5Tnd-2Jj0Mv6co23PW2t3BmHc7eLFj9TgY3eIBg8zislo7xZYNBqovVLcq")},
+			"identifier/cafebabecafebabecafebabecafebabecafebabecafebabecafebabecafebabe": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/0bSFPu5Tnd-2Jj0Mv6co23PW2t3BmHc7eLFj9TgY3eIBg8zislo7xZYNBqovVLcq")},
+			"identifier/deadbeefdeadbeefdeadbeefdeadbeefdeadbeefdeadbeefdeadbeefdeadbeef": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/0bSFPu5Tnd-2Jj0Mv6co23PW2t3BmHc7eLFj9TgY3eIBg8zislo7xZYNBqovVLcq")},
+
+			"work": {Mode: fs.ModeDir | 0700},
+		}, []pkg.FlatEntry{
+			{Mode: fs.ModeDir | 0700, Path: "."},
+
+			{Mode: fs.ModeDir | 0700, Path: "checksum"},
+			{Mode: 0400, Path: "checksum/0bSFPu5Tnd-2Jj0Mv6co23PW2t3BmHc7eLFj9TgY3eIBg8zislo7xZYNBqovVLcq", Data: []byte{0, 0, 0, 0, 0xad, 0xb, 0, 4, 0xfe, 0xfe, 0, 0, 0xfe, 0xca, 0, 0}},
+			{Mode: 0400, Path: "checksum/vsAhtPNo4waRNOASwrQwcIPTqb3SBuJOXw2G4T1mNmVZM-wrQTRllmgXqcIIoRcX", Data: []byte{0}},
+
+			{Mode: fs.ModeDir | 0700, Path: "identifier"},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/0bSFPu5Tnd-2Jj0Mv6co23PW2t3BmHc7eLFj9TgY3eIBg8zislo7xZYNBqovVLcq", Data: []byte("../checksum/0bSFPu5Tnd-2Jj0Mv6co23PW2t3BmHc7eLFj9TgY3eIBg8zislo7xZYNBqovVLcq")},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/cafebabecafebabecafebabecafebabecafebabecafebabecafebabecafebabe", Data: []byte("../checksum/0bSFPu5Tnd-2Jj0Mv6co23PW2t3BmHc7eLFj9TgY3eIBg8zislo7xZYNBqovVLcq")},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/deadbeefdeadbeefdeadbeefdeadbeefdeadbeefdeadbeefdeadbeefdeadbeef", Data: []byte("../checksum/0bSFPu5Tnd-2Jj0Mv6co23PW2t3BmHc7eLFj9TgY3eIBg8zislo7xZYNBqovVLcq")},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/vsAhtPNo4waRNOASwrQwcIPTqb3SBuJOXw2G4T1mNmVZM-wrQTRllmgXqcIIoRcX", Data: []byte("../checksum/vsAhtPNo4waRNOASwrQwcIPTqb3SBuJOXw2G4T1mNmVZM-wrQTRllmgXqcIIoRcX")},
+
+			{Mode: fs.ModeDir | 0700, Path: "work"},
+		}, pkg.MustDecode("St9rlE-mGZ5gXwiv_hzQ_B8bZP-UUvSNmf4nHUZzCMOumb6hKnheZSe0dmnuc4Q2"), nil},
+
+		{"sample http get cure", fstest.MapFS{
+			".": {Mode: fs.ModeDir | 0700},
+
+			"checksum": {Mode: fs.ModeDir | 0700},
+			"checksum/fLYGIMHgN1louE-JzITJZJo2SDniPu-IHBXubtvQWFO-hXnDVKNuscV7-zlyr5fU": {Mode: 0400, Data: []byte("\x7f\xe1\x69\xa2\xdd\x63\x96\x26\x83\x79\x61\x8b\xf0\x3f\xd5\x16\x9a\x39\x3a\xdb\xcf\xb1\xbc\x8d\x33\xff\x75\xee\x62\x56\xa9\xf0\x27\xac\x13\x94\x69")},
+
+			"identifier": {Mode: fs.ModeDir | 0700},
+			"identifier/NqVORkT6L9HX6Za7kT2zcibY10qFqBaxEjPiYFrBQX-ZFr3yxCzJxbKOP0zVjeWb": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/fLYGIMHgN1louE-JzITJZJo2SDniPu-IHBXubtvQWFO-hXnDVKNuscV7-zlyr5fU")},
+
+			"work": {Mode: fs.ModeDir | 0700},
+		}, []pkg.FlatEntry{
+			{Mode: fs.ModeDir | 0700, Path: "."},
+
+			{Mode: fs.ModeDir | 0700, Path: "checksum"},
+			{Mode: 0400, Path: "checksum/fLYGIMHgN1louE-JzITJZJo2SDniPu-IHBXubtvQWFO-hXnDVKNuscV7-zlyr5fU", Data: []byte("\x7f\xe1\x69\xa2\xdd\x63\x96\x26\x83\x79\x61\x8b\xf0\x3f\xd5\x16\x9a\x39\x3a\xdb\xcf\xb1\xbc\x8d\x33\xff\x75\xee\x62\x56\xa9\xf0\x27\xac\x13\x94\x69")},
+
+			{Mode: fs.ModeDir | 0700, Path: "identifier"},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/NqVORkT6L9HX6Za7kT2zcibY10qFqBaxEjPiYFrBQX-ZFr3yxCzJxbKOP0zVjeWb", Data: []byte("../checksum/fLYGIMHgN1louE-JzITJZJo2SDniPu-IHBXubtvQWFO-hXnDVKNuscV7-zlyr5fU")},
+
+			{Mode: fs.ModeDir | 0700, Path: "work"},
+		}, pkg.MustDecode("bqtn69RkV5E7V7GhhgCFjcvbxmaqrO8DywamM4Tyjf10F6EJBHjXiIa_tFRtF4iN"), nil},
+
+		{"sample directory step simple", fstest.MapFS{
+			".": {Mode: fs.ModeDir | 0500},
+
+			"check": {Mode: 0400, Data: []byte{0, 0}},
+
+			"lib":            {Mode: fs.ModeDir | 0700},
+			"lib/libedac.so": {Mode: fs.ModeSymlink | 0777, Data: []byte("/proc/nonexistent/libedac.so")},
+
+			"lib/pkgconfig": {Mode: fs.ModeDir | 0700},
+		}, []pkg.FlatEntry{
+			{Mode: fs.ModeDir | 0500, Path: "."},
+
+			{Mode: 0400, Path: "check", Data: []byte{0, 0}},
+
+			{Mode: fs.ModeDir | 0700, Path: "lib"},
+			{Mode: fs.ModeSymlink | 0777, Path: "lib/libedac.so", Data: []byte("/proc/nonexistent/libedac.so")},
+
+			{Mode: fs.ModeDir | 0700, Path: "lib/pkgconfig"},
+		}, pkg.MustDecode("qRN6in76LndiiOZJheHkwyW8UT1N5-f-bXvHfDvwrMw2fSkOoZdh8pWE1qhLk65b"), nil},
+
+		{"sample directory step garbage", fstest.MapFS{
+			".": {Mode: fs.ModeDir | 0500},
+
+			"lib":       {Mode: fs.ModeDir | 0500},
+			"lib/check": {Mode: 0400, Data: []byte{}},
+
+			"lib/pkgconfig": {Mode: fs.ModeDir | 0500},
+		}, []pkg.FlatEntry{
+			{Mode: fs.ModeDir | 0500, Path: "."},
+
+			{Mode: fs.ModeDir | 0500, Path: "lib"},
+			{Mode: 0400, Path: "lib/check", Data: []byte{}},
+
+			{Mode: fs.ModeDir | 0500, Path: "lib/pkgconfig"},
+		}, pkg.MustDecode("CUx-3hSbTWPsbMfDhgalG4Ni_GmR9TnVX8F99tY_P5GtkYvczg9RrF5zO0jX9XYT"), nil},
+
+		{"sample directory", fstest.MapFS{
+			".": {Mode: fs.ModeDir | 0700},
+
+			"checksum": {Mode: fs.ModeDir | 0700},
+			"checksum/qRN6in76LndiiOZJheHkwyW8UT1N5-f-bXvHfDvwrMw2fSkOoZdh8pWE1qhLk65b":                {Mode: fs.ModeDir | 0500},
+			"checksum/qRN6in76LndiiOZJheHkwyW8UT1N5-f-bXvHfDvwrMw2fSkOoZdh8pWE1qhLk65b/check":          {Mode: 0400, Data: []byte{0, 0}},
+			"checksum/qRN6in76LndiiOZJheHkwyW8UT1N5-f-bXvHfDvwrMw2fSkOoZdh8pWE1qhLk65b/lib":            {Mode: fs.ModeDir | 0700},
+			"checksum/qRN6in76LndiiOZJheHkwyW8UT1N5-f-bXvHfDvwrMw2fSkOoZdh8pWE1qhLk65b/lib/pkgconfig":  {Mode: fs.ModeDir | 0700},
+			"checksum/qRN6in76LndiiOZJheHkwyW8UT1N5-f-bXvHfDvwrMw2fSkOoZdh8pWE1qhLk65b/lib/libedac.so": {Mode: fs.ModeSymlink | 0777, Data: []byte("/proc/nonexistent/libedac.so")},
+
+			"identifier": {Mode: fs.ModeDir | 0700},
+			"identifier/HnySzeLQvSBZuTUcvfmLEX_OmH4yJWWH788NxuLuv7kVn8_uPM6Ks4rqFWM2NZJY": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/qRN6in76LndiiOZJheHkwyW8UT1N5-f-bXvHfDvwrMw2fSkOoZdh8pWE1qhLk65b")},
+			"identifier/Zx5ZG9BAwegNT3zQwCySuI2ktCXxNgxirkGLFjW4FW06PtojYVaCdtEw8yuntPLa": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/qRN6in76LndiiOZJheHkwyW8UT1N5-f-bXvHfDvwrMw2fSkOoZdh8pWE1qhLk65b")},
+
+			"work": {Mode: fs.ModeDir | 0700},
+		}, []pkg.FlatEntry{
+			{Mode: fs.ModeDir | 0700, Path: "."},
+
+			{Mode: fs.ModeDir | 0700, Path: "checksum"},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/qRN6in76LndiiOZJheHkwyW8UT1N5-f-bXvHfDvwrMw2fSkOoZdh8pWE1qhLk65b"},
+			{Mode: 0400, Path: "checksum/qRN6in76LndiiOZJheHkwyW8UT1N5-f-bXvHfDvwrMw2fSkOoZdh8pWE1qhLk65b/check", Data: []byte{0, 0}},
+			{Mode: fs.ModeDir | 0700, Path: "checksum/qRN6in76LndiiOZJheHkwyW8UT1N5-f-bXvHfDvwrMw2fSkOoZdh8pWE1qhLk65b/lib"},
+			{Mode: fs.ModeSymlink | 0777, Path: "checksum/qRN6in76LndiiOZJheHkwyW8UT1N5-f-bXvHfDvwrMw2fSkOoZdh8pWE1qhLk65b/lib/libedac.so", Data: []byte("/proc/nonexistent/libedac.so")},
+			{Mode: fs.ModeDir | 0700, Path: "checksum/qRN6in76LndiiOZJheHkwyW8UT1N5-f-bXvHfDvwrMw2fSkOoZdh8pWE1qhLk65b/lib/pkgconfig"},
+
+			{Mode: fs.ModeDir | 0700, Path: "identifier"},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/HnySzeLQvSBZuTUcvfmLEX_OmH4yJWWH788NxuLuv7kVn8_uPM6Ks4rqFWM2NZJY", Data: []byte("../checksum/qRN6in76LndiiOZJheHkwyW8UT1N5-f-bXvHfDvwrMw2fSkOoZdh8pWE1qhLk65b")},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/Zx5ZG9BAwegNT3zQwCySuI2ktCXxNgxirkGLFjW4FW06PtojYVaCdtEw8yuntPLa", Data: []byte("../checksum/qRN6in76LndiiOZJheHkwyW8UT1N5-f-bXvHfDvwrMw2fSkOoZdh8pWE1qhLk65b")},
+
+			{Mode: fs.ModeDir | 0700, Path: "work"},
+		}, pkg.MustDecode("WVpvsVqVKg9Nsh744x57h51AuWUoUR2nnh8Md-EYBQpk6ziyTuUn6PLtF2e0Eu_d"), nil},
+
+		{"sample tar step unpack", fstest.MapFS{
+			".": {Mode: fs.ModeDir | 0500},
+
+			"checksum": {Mode: fs.ModeDir | 0500},
+			"checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP":                {Mode: fs.ModeDir | 0500},
+			"checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP/check":          {Mode: 0400, Data: []byte{0, 0}},
+			"checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP/lib":            {Mode: fs.ModeDir | 0500},
+			"checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP/lib/pkgconfig":  {Mode: fs.ModeDir | 0500},
+			"checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP/lib/libedac.so": {Mode: fs.ModeSymlink | 0777, Data: []byte("/proc/nonexistent/libedac.so")},
+
+			"identifier": {Mode: fs.ModeDir | 0500},
+			"identifier/HnySzeLQvSBZuTUcvfmLEX_OmH4yJWWH788NxuLuv7kVn8_uPM6Ks4rqFWM2NZJY": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP")},
+			"identifier/Zx5ZG9BAwegNT3zQwCySuI2ktCXxNgxirkGLFjW4FW06PtojYVaCdtEw8yuntPLa": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP")},
+
+			"work": {Mode: fs.ModeDir | 0500},
+		}, []pkg.FlatEntry{
+			{Mode: fs.ModeDir | 0500, Path: "."},
+
+			{Mode: fs.ModeDir | 0500, Path: "checksum"},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP"},
+			{Mode: 0400, Path: "checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP/check", Data: []byte{0, 0}},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP/lib"},
+			{Mode: fs.ModeSymlink | 0777, Path: "checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP/lib/libedac.so", Data: []byte("/proc/nonexistent/libedac.so")},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP/lib/pkgconfig"},
+
+			{Mode: fs.ModeDir | 0500, Path: "identifier"},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/HnySzeLQvSBZuTUcvfmLEX_OmH4yJWWH788NxuLuv7kVn8_uPM6Ks4rqFWM2NZJY", Data: []byte("../checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP")},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/Zx5ZG9BAwegNT3zQwCySuI2ktCXxNgxirkGLFjW4FW06PtojYVaCdtEw8yuntPLa", Data: []byte("../checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP")},
+
+			{Mode: fs.ModeDir | 0500, Path: "work"},
+		}, pkg.MustDecode("cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM"), nil},
+
+		{"sample tar", fstest.MapFS{
+			".": {Mode: fs.ModeDir | 0700},
+
+			"checksum": {Mode: fs.ModeDir | 0700},
+			"checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM":                                                                                          {Mode: fs.ModeDir | 0500},
+			"checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM/checksum":                                                                                 {Mode: fs.ModeDir | 0500},
+			"checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM/checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP":                {Mode: fs.ModeDir | 0500},
+			"checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM/checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP/check":          {Mode: 0400, Data: []byte{0, 0}},
+			"checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM/checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP/lib":            {Mode: fs.ModeDir | 0500},
+			"checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM/checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP/lib/libedac.so": {Mode: fs.ModeSymlink | 0777, Data: []byte("/proc/nonexistent/libedac.so")},
+			"checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM/checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP/lib/pkgconfig":  {Mode: fs.ModeDir | 0500},
+			"checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM/identifier":                                                                               {Mode: fs.ModeDir | 0500},
+			"checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM/identifier/HnySzeLQvSBZuTUcvfmLEX_OmH4yJWWH788NxuLuv7kVn8_uPM6Ks4rqFWM2NZJY":              {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP")},
+			"checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM/identifier/Zx5ZG9BAwegNT3zQwCySuI2ktCXxNgxirkGLFjW4FW06PtojYVaCdtEw8yuntPLa":              {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP")},
+			"checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM/work":                                                                                     {Mode: fs.ModeDir | 0500},
+
+			"identifier": {Mode: fs.ModeDir | 0700},
+			"identifier/-P_1iw6yVq_letMHncqcExSE0bYcDhYI5OdY6b1wKASf-Corufvj__XTBUq2Qd2a": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM")},
+			"identifier/0_rRxIqbX9LK9L_KDbuafotFz6HFkonNgO9gXhK1asM_Y1Pxn0amg756vRTo6m74": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM")},
+
+			"temp": {Mode: fs.ModeDir | 0700},
+			"work": {Mode: fs.ModeDir | 0700},
+		}, []pkg.FlatEntry{
+			{Mode: fs.ModeDir | 0700, Path: "."},
+
+			{Mode: fs.ModeDir | 0700, Path: "checksum"},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM"},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM/checksum"},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM/checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP"},
+			{Mode: 0400, Path: "checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM/checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP/check", Data: []byte{0, 0}},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM/checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP/lib"},
+			{Mode: fs.ModeSymlink | 0777, Path: "checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM/checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP/lib/libedac.so", Data: []byte("/proc/nonexistent/libedac.so")},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM/checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP/lib/pkgconfig"},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM/identifier"},
+			{Mode: fs.ModeSymlink | 0777, Path: "checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM/identifier/HnySzeLQvSBZuTUcvfmLEX_OmH4yJWWH788NxuLuv7kVn8_uPM6Ks4rqFWM2NZJY", Data: []byte("../checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP")},
+			{Mode: fs.ModeSymlink | 0777, Path: "checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM/identifier/Zx5ZG9BAwegNT3zQwCySuI2ktCXxNgxirkGLFjW4FW06PtojYVaCdtEw8yuntPLa", Data: []byte("../checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP")},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM/work"},
+
+			{Mode: fs.ModeDir | 0700, Path: "identifier"},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/-P_1iw6yVq_letMHncqcExSE0bYcDhYI5OdY6b1wKASf-Corufvj__XTBUq2Qd2a", Data: []byte("../checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM")},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/0_rRxIqbX9LK9L_KDbuafotFz6HFkonNgO9gXhK1asM_Y1Pxn0amg756vRTo6m74", Data: []byte("../checksum/cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM")},
+
+			{Mode: fs.ModeDir | 0700, Path: "temp"},
+			{Mode: fs.ModeDir | 0700, Path: "work"},
+		}, pkg.MustDecode("sxbgyX-bPoezbha214n2lbQhiVfTUBkhZ0EX6zI7mmkMdrCdwuMwhMBJphLQsy94"), nil},
+
+		{"sample tar expand step unpack", fstest.MapFS{
+			".": {Mode: fs.ModeDir | 0500},
+
+			"libedac.so": {Mode: fs.ModeSymlink | 0777, Data: []byte("/proc/nonexistent/libedac.so")},
+		}, []pkg.FlatEntry{
+			{Mode: fs.ModeDir | 0500, Path: "."},
+
+			{Mode: fs.ModeSymlink | 0777, Path: "libedac.so", Data: []byte("/proc/nonexistent/libedac.so")},
+		}, pkg.MustDecode("CH3AiUrCCcVOjOYLaMKKK1Da78989JtfHeIsxMzWOQFiN4mrCLDYpoDxLWqJWCUN"), nil},
+
+		{"sample tar expand", fstest.MapFS{
+			".": {Mode: fs.ModeDir | 0700},
+
+			"checksum": {Mode: fs.ModeDir | 0700},
+			"checksum/CH3AiUrCCcVOjOYLaMKKK1Da78989JtfHeIsxMzWOQFiN4mrCLDYpoDxLWqJWCUN":            {Mode: fs.ModeDir | 0500},
+			"checksum/CH3AiUrCCcVOjOYLaMKKK1Da78989JtfHeIsxMzWOQFiN4mrCLDYpoDxLWqJWCUN/libedac.so": {Mode: fs.ModeSymlink | 0777, Data: []byte("/proc/nonexistent/libedac.so")},
+
+			"identifier": {Mode: fs.ModeDir | 0700},
+			"identifier/-P_1iw6yVq_letMHncqcExSE0bYcDhYI5OdY6b1wKASf-Corufvj__XTBUq2Qd2a": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/CH3AiUrCCcVOjOYLaMKKK1Da78989JtfHeIsxMzWOQFiN4mrCLDYpoDxLWqJWCUN")},
+			"identifier/0_rRxIqbX9LK9L_KDbuafotFz6HFkonNgO9gXhK1asM_Y1Pxn0amg756vRTo6m74": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/CH3AiUrCCcVOjOYLaMKKK1Da78989JtfHeIsxMzWOQFiN4mrCLDYpoDxLWqJWCUN")},
+
+			"temp": {Mode: fs.ModeDir | 0700},
+			"work": {Mode: fs.ModeDir | 0700},
+		}, []pkg.FlatEntry{
+			{Mode: fs.ModeDir | 0700, Path: "."},
+
+			{Mode: fs.ModeDir | 0700, Path: "checksum"},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/CH3AiUrCCcVOjOYLaMKKK1Da78989JtfHeIsxMzWOQFiN4mrCLDYpoDxLWqJWCUN"},
+			{Mode: fs.ModeSymlink | 0777, Path: "checksum/CH3AiUrCCcVOjOYLaMKKK1Da78989JtfHeIsxMzWOQFiN4mrCLDYpoDxLWqJWCUN/libedac.so", Data: []byte("/proc/nonexistent/libedac.so")},
+
+			{Mode: fs.ModeDir | 0700, Path: "identifier"},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/-P_1iw6yVq_letMHncqcExSE0bYcDhYI5OdY6b1wKASf-Corufvj__XTBUq2Qd2a", Data: []byte("../checksum/CH3AiUrCCcVOjOYLaMKKK1Da78989JtfHeIsxMzWOQFiN4mrCLDYpoDxLWqJWCUN")},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/0_rRxIqbX9LK9L_KDbuafotFz6HFkonNgO9gXhK1asM_Y1Pxn0amg756vRTo6m74", Data: []byte("../checksum/CH3AiUrCCcVOjOYLaMKKK1Da78989JtfHeIsxMzWOQFiN4mrCLDYpoDxLWqJWCUN")},
+
+			{Mode: fs.ModeDir | 0700, Path: "temp"},
+			{Mode: fs.ModeDir | 0700, Path: "work"},
+		}, pkg.MustDecode("4I8wx_h7NSJTlG5lbuz-GGEXrOg0GYC3M_503LYEBhv5XGWXfNIdIY9Q3eVSYldX"), nil},
+
+		{"testtool", fstest.MapFS{
+			".": {Mode: fs.ModeDir | 0500},
+
+			"check": {Mode: 0400, Data: []byte{0}},
+		}, []pkg.FlatEntry{
+			{Mode: fs.ModeDir | 0500, Path: "."},
+
+			{Mode: 0400, Path: "check", Data: []byte{0}},
+		}, pkg.MustDecode("GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9"), nil},
+
+		{"sample exec container", fstest.MapFS{
+			".": {Mode: fs.ModeDir | 0700},
+
+			"checksum": {Mode: fs.ModeDir | 0700},
+			"checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9":       {Mode: fs.ModeDir | 0500},
+			"checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9/check": {Mode: 0400, Data: []byte{0}},
+			"checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU":       {Mode: fs.ModeDir | 0500},
+			"checksum/OLBgp1GsljhM2TJ-sbHjaiH9txEUvgdDTAzHv2P24donTt6_529l-9Ua0vFImLlb":       {Mode: 0400, Data: []byte{}},
+
+			"identifier": {Mode: fs.ModeDir | 0700},
+			"identifier/U2cbgVgEtjfRuvHfE1cQnZ3t8yoexULQyo_VLgvxAVJSsobMcNaFIsuDWtmt7kzK": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9")},
+			"identifier/_gAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/OLBgp1GsljhM2TJ-sbHjaiH9txEUvgdDTAzHv2P24donTt6_529l-9Ua0vFImLlb")},
+			"identifier/nfeISfLeFDr1k-g3hpE1oZ440kTqDdfF8TDpoLdbTPqaMMIl95oiqcvqjRkMjubA": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU")},
+
+			"temp": {Mode: fs.ModeDir | 0700},
+			"work": {Mode: fs.ModeDir | 0700},
+		}, []pkg.FlatEntry{
+			{Mode: fs.ModeDir | 0700, Path: "."},
+
+			{Mode: fs.ModeDir | 0700, Path: "checksum"},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9"},
+			{Mode: 0400, Path: "checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9/check", Data: []byte{0}},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU"},
+			{Mode: 0400, Path: "checksum/OLBgp1GsljhM2TJ-sbHjaiH9txEUvgdDTAzHv2P24donTt6_529l-9Ua0vFImLlb", Data: []byte{}},
+
+			{Mode: fs.ModeDir | 0700, Path: "identifier"},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/U2cbgVgEtjfRuvHfE1cQnZ3t8yoexULQyo_VLgvxAVJSsobMcNaFIsuDWtmt7kzK", Data: []byte("../checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9")},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/_gAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA", Data: []byte("../checksum/OLBgp1GsljhM2TJ-sbHjaiH9txEUvgdDTAzHv2P24donTt6_529l-9Ua0vFImLlb")},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/nfeISfLeFDr1k-g3hpE1oZ440kTqDdfF8TDpoLdbTPqaMMIl95oiqcvqjRkMjubA", Data: []byte("../checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU")},
+
+			{Mode: fs.ModeDir | 0700, Path: "temp"},
+			{Mode: fs.ModeDir | 0700, Path: "work"},
+		}, pkg.MustDecode("UiV6kMz7KrTsc_yphiyQzFLqjRanHxUOwrBMtkKuWo4mOO6WgPFAcoUEeSp7eVIW"), nil},
+
+		{"testtool net", fstest.MapFS{
+			".": {Mode: fs.ModeDir | 0500},
+
+			"check": {Mode: 0400, Data: []byte("net")},
+		}, []pkg.FlatEntry{
+			{Mode: fs.ModeDir | 0500, Path: "."},
+
+			{Mode: 0400, Path: "check", Data: []byte("net")},
+		}, pkg.MustDecode("a1F_i9PVQI4qMcoHgTQkORuyWLkC1GLIxOhDt2JpU1NGAxWc5VJzdlfRK-PYBh3W"), nil},
+
+		{"sample exec net container", fstest.MapFS{
+			".": {Mode: fs.ModeDir | 0700},
+
+			"checksum": {Mode: fs.ModeDir | 0700},
+			"checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU":       {Mode: fs.ModeDir | 0500},
+			"checksum/OLBgp1GsljhM2TJ-sbHjaiH9txEUvgdDTAzHv2P24donTt6_529l-9Ua0vFImLlb":       {Mode: 0400, Data: []byte{}},
+			"checksum/a1F_i9PVQI4qMcoHgTQkORuyWLkC1GLIxOhDt2JpU1NGAxWc5VJzdlfRK-PYBh3W":       {Mode: fs.ModeDir | 0500},
+			"checksum/a1F_i9PVQI4qMcoHgTQkORuyWLkC1GLIxOhDt2JpU1NGAxWc5VJzdlfRK-PYBh3W/check": {Mode: 0400, Data: []byte("net")},
+
+			"identifier": {Mode: fs.ModeDir | 0700},
+			"identifier/_gAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/OLBgp1GsljhM2TJ-sbHjaiH9txEUvgdDTAzHv2P24donTt6_529l-9Ua0vFImLlb")},
+			"identifier/QdsJhGgnk5N2xdUNGcndXQxFKifxf1V_2t9X8CQ-pDcg24x6mGJC_BiLfGbs6Qml": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/a1F_i9PVQI4qMcoHgTQkORuyWLkC1GLIxOhDt2JpU1NGAxWc5VJzdlfRK-PYBh3W")},
+			"identifier/nfeISfLeFDr1k-g3hpE1oZ440kTqDdfF8TDpoLdbTPqaMMIl95oiqcvqjRkMjubA": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU")},
+
+			"temp": {Mode: fs.ModeDir | 0700},
+			"work": {Mode: fs.ModeDir | 0700},
+		}, []pkg.FlatEntry{
+			{Mode: fs.ModeDir | 0700, Path: "."},
+
+			{Mode: fs.ModeDir | 0700, Path: "checksum"},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU"},
+			{Mode: 0400, Path: "checksum/OLBgp1GsljhM2TJ-sbHjaiH9txEUvgdDTAzHv2P24donTt6_529l-9Ua0vFImLlb", Data: []byte{}},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/a1F_i9PVQI4qMcoHgTQkORuyWLkC1GLIxOhDt2JpU1NGAxWc5VJzdlfRK-PYBh3W"},
+			{Mode: 0400, Path: "checksum/a1F_i9PVQI4qMcoHgTQkORuyWLkC1GLIxOhDt2JpU1NGAxWc5VJzdlfRK-PYBh3W/check", Data: []byte("net")},
+
+			{Mode: fs.ModeDir | 0700, Path: "identifier"},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/QdsJhGgnk5N2xdUNGcndXQxFKifxf1V_2t9X8CQ-pDcg24x6mGJC_BiLfGbs6Qml", Data: []byte("../checksum/a1F_i9PVQI4qMcoHgTQkORuyWLkC1GLIxOhDt2JpU1NGAxWc5VJzdlfRK-PYBh3W")},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/_gAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA", Data: []byte("../checksum/OLBgp1GsljhM2TJ-sbHjaiH9txEUvgdDTAzHv2P24donTt6_529l-9Ua0vFImLlb")},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/nfeISfLeFDr1k-g3hpE1oZ440kTqDdfF8TDpoLdbTPqaMMIl95oiqcvqjRkMjubA", Data: []byte("../checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU")},
+
+			{Mode: fs.ModeDir | 0700, Path: "temp"},
+			{Mode: fs.ModeDir | 0700, Path: "work"},
+		}, pkg.MustDecode("ek4K-0d4iRSArkY2TCs3WK34DbiYeOmhE_4vsJTSu_6roY4ZF3YG6eKRooal-i1o"), nil},
+
+		{"sample exec container overlay root", fstest.MapFS{
+			".": {Mode: fs.ModeDir | 0700},
+
+			"checksum": {Mode: fs.ModeDir | 0700},
+			"checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9":       {Mode: fs.ModeDir | 0500},
+			"checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9/check": {Mode: 0400, Data: []byte{0}},
+			"checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU":       {Mode: fs.ModeDir | 0500},
+
+			"identifier": {Mode: fs.ModeDir | 0700},
+			"identifier/5ey2wpmMpj483YYa7ZZQciYLA2cx3_l167JCqWW4Pd-5DVp81dj9EsBtVTwYptF6": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9")},
+			"identifier/nfeISfLeFDr1k-g3hpE1oZ440kTqDdfF8TDpoLdbTPqaMMIl95oiqcvqjRkMjubA": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU")},
+
+			"temp": {Mode: fs.ModeDir | 0700},
+			"work": {Mode: fs.ModeDir | 0700},
+		}, []pkg.FlatEntry{
+			{Mode: fs.ModeDir | 0700, Path: "."},
+
+			{Mode: fs.ModeDir | 0700, Path: "checksum"},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9"},
+			{Mode: 0400, Path: "checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9/check", Data: []byte{0}},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU"},
+
+			{Mode: fs.ModeDir | 0700, Path: "identifier"},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/5ey2wpmMpj483YYa7ZZQciYLA2cx3_l167JCqWW4Pd-5DVp81dj9EsBtVTwYptF6", Data: []byte("../checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9")},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/nfeISfLeFDr1k-g3hpE1oZ440kTqDdfF8TDpoLdbTPqaMMIl95oiqcvqjRkMjubA", Data: []byte("../checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU")},
+
+			{Mode: fs.ModeDir | 0700, Path: "temp"},
+			{Mode: fs.ModeDir | 0700, Path: "work"},
+		}, pkg.MustDecode("VIqqpf0ip9jcyw63i6E8lCMGUcLivQBe4Bevt3WusNac-1MSy5bzB647qGUBzl-W"), nil},
+
+		{"sample exec container overlay work", fstest.MapFS{
+			".": {Mode: fs.ModeDir | 0700},
+
+			"checksum": {Mode: fs.ModeDir | 0700},
+			"checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9":       {Mode: fs.ModeDir | 0500},
+			"checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9/check": {Mode: 0400, Data: []byte{0}},
+			"checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU":       {Mode: fs.ModeDir | 0500},
+
+			"identifier": {Mode: fs.ModeDir | 0700},
+			"identifier/acaDzHZv40dZaz4cGAXayqbRMgbEOuiuiUijZL8IgDQvyeCNMFE3onBMYfny-kXA": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9")},
+			"identifier/nfeISfLeFDr1k-g3hpE1oZ440kTqDdfF8TDpoLdbTPqaMMIl95oiqcvqjRkMjubA": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU")},
+
+			"temp": {Mode: fs.ModeDir | 0700},
+			"work": {Mode: fs.ModeDir | 0700},
+		}, []pkg.FlatEntry{
+			{Mode: fs.ModeDir | 0700, Path: "."},
+
+			{Mode: fs.ModeDir | 0700, Path: "checksum"},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9"},
+			{Mode: 0400, Path: "checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9/check", Data: []byte{0}},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU"},
+
+			{Mode: fs.ModeDir | 0700, Path: "identifier"},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/acaDzHZv40dZaz4cGAXayqbRMgbEOuiuiUijZL8IgDQvyeCNMFE3onBMYfny-kXA", Data: []byte("../checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9")},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/nfeISfLeFDr1k-g3hpE1oZ440kTqDdfF8TDpoLdbTPqaMMIl95oiqcvqjRkMjubA", Data: []byte("../checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU")},
+
+			{Mode: fs.ModeDir | 0700, Path: "temp"},
+			{Mode: fs.ModeDir | 0700, Path: "work"},
+		}, pkg.MustDecode("q8x2zQg4YZbKpPqKlEBj_uxXD9vOBaZ852qOuIsl9QdO73I_UMNpuUoPLtunxUYl"), nil},
+
+		{"sample exec container multiple layers", fstest.MapFS{
+			".": {Mode: fs.ModeDir | 0700},
+
+			"checksum": {Mode: fs.ModeDir | 0700},
+			"checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9":       {Mode: fs.ModeDir | 0500},
+			"checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9/check": {Mode: 0400, Data: []byte{0}},
+			"checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU":       {Mode: fs.ModeDir | 0500},
+			"checksum/OLBgp1GsljhM2TJ-sbHjaiH9txEUvgdDTAzHv2P24donTt6_529l-9Ua0vFImLlb":       {Mode: 0400, Data: []byte{}},
+			"checksum/nY_CUdiaUM1OL4cPr5TS92FCJ3rCRV7Hm5oVTzAvMXwC03_QnTRfQ5PPs7mOU9fK":       {Mode: fs.ModeDir | 0500},
+			"checksum/nY_CUdiaUM1OL4cPr5TS92FCJ3rCRV7Hm5oVTzAvMXwC03_QnTRfQ5PPs7mOU9fK/check": {Mode: 0400, Data: []byte("layers")},
+
+			"identifier": {Mode: fs.ModeDir | 0700},
+			"identifier/_gAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/OLBgp1GsljhM2TJ-sbHjaiH9txEUvgdDTAzHv2P24donTt6_529l-9Ua0vFImLlb")},
+			"identifier/nfeISfLeFDr1k-g3hpE1oZ440kTqDdfF8TDpoLdbTPqaMMIl95oiqcvqjRkMjubA": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU")},
+			"identifier/rXLKjjYfGSyoWmuvEJooHkvGJIZaC0IAWnKGvtPZkM15gBxAgW7mIXcxRVNOXAr4": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/nY_CUdiaUM1OL4cPr5TS92FCJ3rCRV7Hm5oVTzAvMXwC03_QnTRfQ5PPs7mOU9fK")},
+			"identifier/tfjrsVuBuFgzWgwz-yPppFtylYuC1VFWnKhyBiHbWTGkyz8lt7Ee9QXWaIHPXs4x": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9")},
+
+			"temp": {Mode: fs.ModeDir | 0700},
+			"work": {Mode: fs.ModeDir | 0700},
+		}, []pkg.FlatEntry{
+			{Mode: fs.ModeDir | 0700, Path: "."},
+
+			{Mode: fs.ModeDir | 0700, Path: "checksum"},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9"},
+			{Mode: 0400, Path: "checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9/check", Data: []byte{0}},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU"},
+			{Mode: 0400, Path: "checksum/OLBgp1GsljhM2TJ-sbHjaiH9txEUvgdDTAzHv2P24donTt6_529l-9Ua0vFImLlb", Data: []byte{}},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/nY_CUdiaUM1OL4cPr5TS92FCJ3rCRV7Hm5oVTzAvMXwC03_QnTRfQ5PPs7mOU9fK"},
+			{Mode: 0400, Path: "checksum/nY_CUdiaUM1OL4cPr5TS92FCJ3rCRV7Hm5oVTzAvMXwC03_QnTRfQ5PPs7mOU9fK/check", Data: []byte("layers")},
+
+			{Mode: fs.ModeDir | 0700, Path: "identifier"},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/_gAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA", Data: []byte("../checksum/OLBgp1GsljhM2TJ-sbHjaiH9txEUvgdDTAzHv2P24donTt6_529l-9Ua0vFImLlb")},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/nfeISfLeFDr1k-g3hpE1oZ440kTqDdfF8TDpoLdbTPqaMMIl95oiqcvqjRkMjubA", Data: []byte("../checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU")},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/rXLKjjYfGSyoWmuvEJooHkvGJIZaC0IAWnKGvtPZkM15gBxAgW7mIXcxRVNOXAr4", Data: []byte("../checksum/nY_CUdiaUM1OL4cPr5TS92FCJ3rCRV7Hm5oVTzAvMXwC03_QnTRfQ5PPs7mOU9fK")},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/tfjrsVuBuFgzWgwz-yPppFtylYuC1VFWnKhyBiHbWTGkyz8lt7Ee9QXWaIHPXs4x", Data: []byte("../checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9")},
+
+			{Mode: fs.ModeDir | 0700, Path: "temp"},
+			{Mode: fs.ModeDir | 0700, Path: "work"},
+		}, pkg.MustDecode("SITnQ6PTV12PAQQjIuLUxkvsXQiC9Gq_HJQlcb4BPL5YnRHnx8lsW7PRM9YMLBsx"), nil},
+
+		{"sample exec container layer promotion", fstest.MapFS{
+			".": {Mode: fs.ModeDir | 0700},
+
+			"checksum": {Mode: fs.ModeDir | 0700},
+			"checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9":       {Mode: fs.ModeDir | 0500},
+			"checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9/check": {Mode: 0400, Data: []byte{0}},
+			"checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU":       {Mode: fs.ModeDir | 0500},
+
+			"identifier": {Mode: fs.ModeDir | 0700},
+			"identifier/1tQZOGmVk_JkpyiG84AKW_BXmlK_MvHUbh5WtMuthGbHUq7i7nL1bvdF-LoJbqNh": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU")},
+			"identifier/O-6VjlIUxc4PYLf5v35uhIeL8kkYCbHYklqlmDjFPXe0m4j6GkUDg5qwTzBRESnf": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9")},
+			"identifier/nfeISfLeFDr1k-g3hpE1oZ440kTqDdfF8TDpoLdbTPqaMMIl95oiqcvqjRkMjubA": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU")},
+
+			"temp": {Mode: fs.ModeDir | 0700},
+			"work": {Mode: fs.ModeDir | 0700},
+		}, []pkg.FlatEntry{
+			{Mode: fs.ModeDir | 0700, Path: "."},
+
+			{Mode: fs.ModeDir | 0700, Path: "checksum"},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9"},
+			{Mode: 0400, Path: "checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9/check", Data: []byte{0}},
+			{Mode: fs.ModeDir | 0500, Path: "checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU"},
+
+			{Mode: fs.ModeDir | 0700, Path: "identifier"},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/1tQZOGmVk_JkpyiG84AKW_BXmlK_MvHUbh5WtMuthGbHUq7i7nL1bvdF-LoJbqNh", Data: []byte("../checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU")},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/O-6VjlIUxc4PYLf5v35uhIeL8kkYCbHYklqlmDjFPXe0m4j6GkUDg5qwTzBRESnf", Data: []byte("../checksum/GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9")},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/nfeISfLeFDr1k-g3hpE1oZ440kTqDdfF8TDpoLdbTPqaMMIl95oiqcvqjRkMjubA", Data: []byte("../checksum/MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU")},
+
+			{Mode: fs.ModeDir | 0700, Path: "temp"},
+			{Mode: fs.ModeDir | 0700, Path: "work"},
+		}, pkg.MustDecode("fuC20BhMKr86TYzNPP2A-9P7mGLvdcOiG10exlhRvZm8ySI7csf0LhW3im_26l1N"), nil},
+
+		{"sample file short", fstest.MapFS{
+			".": {Mode: fs.ModeDir | 0700},
+
+			"checksum": {Mode: fs.ModeDir | 0700},
+			"checksum/vsAhtPNo4waRNOASwrQwcIPTqb3SBuJOXw2G4T1mNmVZM-wrQTRllmgXqcIIoRcX": {Mode: 0400, Data: []byte{0}},
+
+			"identifier": {Mode: fs.ModeDir | 0700},
+			"identifier/lIx_W4M7tVOcQ8jh08EJOfXf4brRmkEEjvUa7c17vVUzlmtUxlhhrgqmc9aZhjbn": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/vsAhtPNo4waRNOASwrQwcIPTqb3SBuJOXw2G4T1mNmVZM-wrQTRllmgXqcIIoRcX")},
+
+			"work": {Mode: fs.ModeDir | 0700},
+		}, []pkg.FlatEntry{
+			{Mode: fs.ModeDir | 0700, Path: "."},
+			{Mode: fs.ModeDir | 0700, Path: "checksum"},
+			{Mode: 0400, Path: "checksum/vsAhtPNo4waRNOASwrQwcIPTqb3SBuJOXw2G4T1mNmVZM-wrQTRllmgXqcIIoRcX", Data: []byte{0}},
+
+			{Mode: fs.ModeDir | 0700, Path: "identifier"},
+			{Mode: fs.ModeSymlink | 0777, Path: "identifier/lIx_W4M7tVOcQ8jh08EJOfXf4brRmkEEjvUa7c17vVUzlmtUxlhhrgqmc9aZhjbn", Data: []byte("../checksum/vsAhtPNo4waRNOASwrQwcIPTqb3SBuJOXw2G4T1mNmVZM-wrQTRllmgXqcIIoRcX")},
+
+			{Mode: fs.ModeDir | 0700, Path: "work"},
+		}, pkg.MustDecode("hnrfmJtivNKcgtETsKnU9gP_OwPgpNY3DSUJnmxnmeOODSO-YBvEBiTgieY4AAd7"), nil},
+	}
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Parallel()
+
+			t.Run("roundtrip", func(t *testing.T) {
+				t.Parallel()
+
+				var buf bytes.Buffer
+				if _, err := pkg.Flatten(
+					tc.fsys,
+					".",
+					&buf,
+				); !reflect.DeepEqual(err, tc.err) {
+					t.Fatalf("Flatten: error = %v, want %v", err, tc.err)
+				} else if tc.err != nil {
+					return
+				}
+
+				s := pkg.NewDirScanner(bytes.NewReader(buf.Bytes()), true)
+				var got []pkg.FlatEntry
+				for s.Scan() {
+					got = append(got, *s.Entry())
+				}
+				if err := s.Err(); err != nil {
+					t.Fatalf("Err: error = %v", err)
+				}
+
+				if !reflect.DeepEqual(got, tc.entries) {
+					t.Fatalf("Scan: %#v, want %#v", got, tc.entries)
+				}
+			})
+
+			if tc.err != nil {
+				return
+			}
+
+			t.Run("hash", func(t *testing.T) {
+				t.Parallel()
+
+				var got pkg.Checksum
+				if err := pkg.HashFS(&got, tc.fsys, "."); err != nil {
+					t.Fatalf("HashFS: error = %v", err)
+				} else if got != tc.sum {
+					t.Fatalf("HashFS: %v", &pkg.ChecksumMismatchError{
+						Got:  got,
+						Want: tc.sum,
+					})
+				}
+			})
+		})
+	}
+}

internal/pkg/exec.go

@@ -0,0 +1,411 @@
+package pkg
+
+import (
+	"bufio"
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"os"
+	"os/exec"
+	"path"
+	"slices"
+	"strconv"
+	"syscall"
+	"time"
+	"unique"
+
+	"hakurei.app/container"
+	"hakurei.app/container/check"
+	"hakurei.app/container/fhs"
+	"hakurei.app/container/std"
+	"hakurei.app/message"
+)
+
+// AbsWork is the container pathname [CureContext.GetWorkDir] is mounted on.
+var AbsWork = fhs.AbsRoot.Append("work/")
+
+// ExecPath is a slice of [Artifact] and the [check.Absolute] pathname to make
+// it available at under in the container.
+type ExecPath struct {
+	// Pathname in the container mount namespace.
+	P *check.Absolute
+	// Artifacts to mount on the pathname, must contain at least one [Artifact].
+	// If there are multiple entries or W is true, P is set up as an overlay
+	// mount, and entries of A must not implement [FileArtifact].
+	A []Artifact
+	// Whether to make the mount point writable via the temp directory.
+	W bool
+}
+
+// layers returns pathnames collected from A deduplicated by checksum.
+func (p *ExecPath) layers(f *FContext) []*check.Absolute {
+	msg := f.GetMessage()
+
+	layers := make([]*check.Absolute, 0, len(p.A))
+	checksums := make(map[unique.Handle[Checksum]]struct{}, len(p.A))
+	for i := range p.A {
+		d := p.A[len(p.A)-1-i]
+		pathname, checksum := f.GetArtifact(d)
+		if _, ok := checksums[checksum]; ok {
+			if msg.IsVerbose() {
+				msg.Verbosef(
+					"promoted layer %d as %s",
+					len(p.A)-1-i, reportName(d, f.cache.Ident(d)),
+				)
+			}
+			continue
+		}
+		checksums[checksum] = struct{}{}
+		layers = append(layers, pathname)
+	}
+	slices.Reverse(layers)
+	return layers
+}
+
+// Path returns a populated [ExecPath].
+func Path(pathname *check.Absolute, writable bool, a ...Artifact) ExecPath {
+	return ExecPath{pathname, a, writable}
+}
+
+// MustPath is like [Path], but takes a string pathname via [check.MustAbs].
+func MustPath(pathname string, writable bool, a ...Artifact) ExecPath {
+	return ExecPath{check.MustAbs(pathname), a, writable}
+}
+
+const (
+	// ExecTimeoutDefault replaces out of range [NewExec] timeout values.
+	ExecTimeoutDefault = 15 * time.Minute
+	// ExecTimeoutMax is the arbitrary upper bound of [NewExec] timeout.
+	ExecTimeoutMax = 48 * time.Hour
+)
+
+// An execArtifact is an [Artifact] that produces output by running a program
+// part of another [Artifact] in a [container] to produce its output.
+//
+// Methods of execArtifact does not modify any struct field or underlying arrays
+// referred to by slices.
+type execArtifact struct {
+	// Caller-supplied user-facing reporting name, guaranteed to be nonzero
+	// during initialisation.
+	name string
+	// Caller-supplied inner mount points.
+	paths []ExecPath
+
+	// Passed through to [container.Params].
+	dir *check.Absolute
+	// Passed through to [container.Params].
+	env []string
+	// Passed through to [container.Params].
+	path *check.Absolute
+	// Passed through to [container.Params].
+	args []string
+
+	// Duration the initial process is allowed to run. The zero value is
+	// equivalent to execTimeoutDefault. This value is never encoded in Params
+	// because it cannot affect outcome.
+	timeout time.Duration
+
+	// Caller-supplied exclusivity value, returned as is by IsExclusive.
+	exclusive bool
+}
+
+var _ fmt.Stringer = new(execArtifact)
+
+// execNetArtifact is like execArtifact but implements [KnownChecksum] and has
+// its resulting container keep the host net namespace.
+type execNetArtifact struct {
+	checksum Checksum
+
+	execArtifact
+}
+
+var _ KnownChecksum = new(execNetArtifact)
+
+// Checksum returns the caller-supplied checksum.
+func (a *execNetArtifact) Checksum() Checksum { return a.checksum }
+
+// Kind returns the hardcoded [Kind] constant.
+func (*execNetArtifact) Kind() Kind { return KindExecNet }
+
+// Params is [Checksum] concatenated with [KindExec] params.
+func (a *execNetArtifact) Params(ctx *IContext) {
+	ctx.GetHash().Write(a.checksum[:])
+	a.execArtifact.Params(ctx)
+}
+
+// Cure cures the [Artifact] in the container described by the caller. The
+// container retains host networking.
+func (a *execNetArtifact) Cure(f *FContext) error {
+	return a.cure(f, true)
+}
+
+// NewExec returns a new [Artifact] that executes the program path in a
+// container with specified paths bind mounted read-only in order. A private
+// instance of /proc and /dev is made available to the container.
+//
+// The working and temporary directories are both created and mounted writable
+// on [AbsWork] and [fhs.AbsTmp] respectively. If one or more paths target
+// [AbsWork], the final entry is set up as a writable overlay mount on /work for
+// which the upperdir is the host side work directory. In this configuration,
+// the W field is ignored, and the program must avoid causing whiteout files to
+// be created. Cure fails if upperdir ends up with entries other than directory,
+// regular or symlink.
+//
+// If checksum is non-nil, the resulting [Artifact] implements [KnownChecksum]
+// and its container runs in the host net namespace.
+//
+// The container is allowed to run for the specified duration before the initial
+// process and all processes originating from it is terminated. A zero or
+// negative timeout value is equivalent tp [ExecTimeoutDefault], a timeout value
+// greater than [ExecTimeoutMax] is equivalent to [ExecTimeoutMax].
+//
+// The user-facing name and exclusivity value are not accessible from the
+// container and does not affect curing outcome. Because of this, it is omitted
+// from parameter data for computing identifier.
+func NewExec(
+	name string,
+	checksum *Checksum,
+	timeout time.Duration,
+	exclusive bool,
+
+	dir *check.Absolute,
+	env []string,
+	pathname *check.Absolute,
+	args []string,
+
+	paths ...ExecPath,
+) Artifact {
+	if name == "" {
+		name = "exec-" + path.Base(pathname.String())
+	}
+	if timeout <= 0 {
+		timeout = ExecTimeoutDefault
+	}
+	if timeout > ExecTimeoutMax {
+		timeout = ExecTimeoutMax
+	}
+	a := execArtifact{name, paths, dir, env, pathname, args, timeout, exclusive}
+	if checksum == nil {
+		return &a
+	}
+	return &execNetArtifact{*checksum, a}
+}
+
+// Kind returns the hardcoded [Kind] constant.
+func (*execArtifact) Kind() Kind { return KindExec }
+
+// Params writes paths, executable pathname and args.
+func (a *execArtifact) Params(ctx *IContext) {
+	h := ctx.GetHash()
+
+	_0, _1 := []byte{0}, []byte{1}
+	for _, p := range a.paths {
+		if p.W {
+			h.Write(_1)
+		} else {
+			h.Write(_0)
+		}
+		if p.P != nil {
+			h.Write([]byte(p.P.String()))
+		} else {
+			h.Write([]byte("invalid P\x00"))
+		}
+		h.Write(_0)
+		for _, d := range p.A {
+			ctx.WriteIdent(d)
+		}
+		h.Write(_0)
+	}
+	h.Write(_0)
+	h.Write([]byte(a.dir.String()))
+	h.Write(_0)
+	for _, e := range a.env {
+		h.Write([]byte(e))
+	}
+	h.Write(_0)
+	h.Write([]byte(a.path.String()))
+	h.Write(_0)
+	for _, arg := range a.args {
+		h.Write([]byte(arg))
+	}
+}
+
+// Dependencies returns a slice of all artifacts collected from caller-supplied
+// [ExecPath].
+func (a *execArtifact) Dependencies() []Artifact {
+	artifacts := make([][]Artifact, 0, len(a.paths))
+	for _, p := range a.paths {
+		artifacts = append(artifacts, p.A)
+	}
+	return slices.Concat(artifacts...)
+}
+
+// IsExclusive returns the caller-supplied exclusivity value.
+func (a *execArtifact) IsExclusive() bool { return a.exclusive }
+
+// String returns the caller-supplied reporting name.
+func (a *execArtifact) String() string { return a.name }
+
+// Cure cures the [Artifact] in the container described by the caller.
+func (a *execArtifact) Cure(f *FContext) (err error) {
+	return a.cure(f, false)
+}
+
+const (
+	// execWaitDelay is passed through to [container.Params].
+	execWaitDelay = time.Nanosecond
+)
+
+// scanVerbose prefixes program output for a verbose [message.Msg].
+func scanVerbose(
+	msg message.Msg,
+	done chan<- struct{},
+	prefix string,
+	r io.Reader,
+) {
+	defer close(done)
+	s := bufio.NewScanner(r)
+	s.Buffer(
+		make([]byte, bufio.MaxScanTokenSize),
+		bufio.MaxScanTokenSize<<12,
+	)
+	for s.Scan() {
+		msg.Verbose(prefix, s.Text())
+	}
+	if err := s.Err(); err != nil && !errors.Is(err, os.ErrClosed) {
+		msg.Verbose("*"+prefix, err)
+	}
+}
+
+// cure is like Cure but allows optional host net namespace. This is used for
+// the [KnownChecksum] variant where networking is allowed.
+func (a *execArtifact) cure(f *FContext, hostNet bool) (err error) {
+	overlayWorkIndex := -1
+	for i, p := range a.paths {
+		if p.P == nil || len(p.A) == 0 {
+			return os.ErrInvalid
+		}
+		if p.P.Is(AbsWork) {
+			overlayWorkIndex = i
+		}
+	}
+
+	var artifactCount int
+	for _, p := range a.paths {
+		artifactCount += len(p.A)
+	}
+
+	ctx, cancel := context.WithTimeout(f.Unwrap(), a.timeout)
+	defer cancel()
+
+	z := container.New(ctx, f.GetMessage())
+	z.WaitDelay = execWaitDelay
+	z.SeccompPresets |= std.PresetStrict & ^std.PresetDenyNS
+	z.ParentPerm = 0700
+	z.HostNet = hostNet
+	z.Hostname = "cure"
+	if z.HostNet {
+		z.Hostname = "cure-net"
+	}
+	z.Uid, z.Gid = (1<<10)-1, (1<<10)-1
+	if msg := f.GetMessage(); msg.IsVerbose() {
+		var stdout, stderr io.ReadCloser
+		if stdout, err = z.StdoutPipe(); err != nil {
+			return
+		}
+		if stderr, err = z.StderrPipe(); err != nil {
+			_ = stdout.Close()
+			return
+		}
+		defer func() {
+			if err != nil && !errors.As(err, new(*exec.ExitError)) {
+				_ = stdout.Close()
+				_ = stderr.Close()
+			}
+		}()
+
+		stdoutDone, stderrDone := make(chan struct{}), make(chan struct{})
+		go scanVerbose(msg, stdoutDone, "("+a.name+":1)", stdout)
+		go scanVerbose(msg, stderrDone, "("+a.name+":2)", stderr)
+		defer func() { <-stdoutDone; <-stderrDone }()
+	}
+
+	z.Dir, z.Env, z.Path, z.Args = a.dir, a.env, a.path, a.args
+	z.Grow(len(a.paths) + 4)
+
+	temp, work := f.GetTempDir(), f.GetWorkDir()
+	for i, b := range a.paths {
+		if i == overlayWorkIndex {
+			if err = os.MkdirAll(work.String(), 0700); err != nil {
+				return
+			}
+			tempWork := temp.Append(".work")
+			if err = os.MkdirAll(tempWork.String(), 0700); err != nil {
+				return
+			}
+			z.Overlay(
+				AbsWork,
+				work,
+				tempWork,
+				b.layers(f)...,
+			)
+			continue
+		}
+
+		if a.paths[i].W {
+			tempUpper, tempWork := temp.Append(
+				".upper", strconv.Itoa(i),
+			), temp.Append(
+				".work", strconv.Itoa(i),
+			)
+			if err = os.MkdirAll(tempUpper.String(), 0700); err != nil {
+				return
+			}
+			if err = os.MkdirAll(tempWork.String(), 0700); err != nil {
+				return
+			}
+			z.Overlay(b.P, tempUpper, tempWork, b.layers(f)...)
+		} else if len(b.A) == 1 {
+			pathname, _ := f.GetArtifact(b.A[0])
+			z.Bind(pathname, b.P, 0)
+		} else {
+			z.OverlayReadonly(b.P, b.layers(f)...)
+		}
+	}
+	if overlayWorkIndex < 0 {
+		z.Bind(
+			work,
+			AbsWork,
+			std.BindWritable|std.BindEnsure,
+		)
+	}
+	z.Bind(
+		f.GetTempDir(),
+		fhs.AbsTmp,
+		std.BindWritable|std.BindEnsure,
+	)
+	z.Proc(fhs.AbsProc).Dev(fhs.AbsDev, true)
+
+	if err = z.Start(); err != nil {
+		return
+	}
+	if err = z.Serve(); err != nil {
+		return
+	}
+	if err = z.Wait(); err != nil {
+		return
+	}
+
+	// do not allow empty directories to succeed
+	for {
+		err = syscall.Rmdir(work.String())
+		if err != syscall.EINTR {
+			break
+		}
+	}
+	if err != nil && errors.Is(err, syscall.ENOTEMPTY) {
+		err = nil
+	}
+	return
+}

internal/pkg/exec_test.go

@@ -0,0 +1,339 @@
+package pkg_test
+
+//go:generate env CGO_ENABLED=0 go build -tags testtool -o testdata/testtool ./testdata
+
+import (
+	_ "embed"
+	"encoding/gob"
+	"errors"
+	"net"
+	"os"
+	"os/exec"
+	"slices"
+	"testing"
+	"unique"
+
+	"hakurei.app/container/check"
+	"hakurei.app/container/stub"
+	"hakurei.app/hst"
+	"hakurei.app/internal/pkg"
+)
+
+// testtoolBin is the container test tool binary made available to the
+// execArtifact for testing its curing environment.
+//
+//go:embed testdata/testtool
+var testtoolBin []byte
+
+func TestExec(t *testing.T) {
+	t.Parallel()
+
+	wantChecksumOffline := pkg.MustDecode(
+		"GPa4aBakdSJd7Tz7LYj_VJFoojzyZinmVcG3k6M5xI6CZ821J5sXLhLDDuS47gi9",
+	)
+
+	checkWithCache(t, []cacheTestCase{
+		{"offline", nil, func(t *testing.T, base *check.Absolute, c *pkg.Cache) {
+			c.SetStrict(true)
+			testtool, testtoolDestroy := newTesttool()
+
+			cureMany(t, c, []cureStep{
+				{"container", pkg.NewExec(
+					"exec-offline", nil, 0, false,
+					pkg.AbsWork,
+					[]string{"HAKUREI_TEST=1"},
+					check.MustAbs("/opt/bin/testtool"),
+					[]string{"testtool"},
+
+					pkg.MustPath("/file", false, newStubFile(
+						pkg.KindHTTPGet,
+						pkg.ID{0xfe, 0},
+						nil,
+						nil, nil,
+					)),
+					pkg.MustPath("/.hakurei", false, &stubArtifact{
+						kind:   pkg.KindTar,
+						params: []byte("empty directory"),
+						cure: func(t *pkg.TContext) error {
+							return os.MkdirAll(t.GetWorkDir().String(), 0700)
+						},
+					}),
+					pkg.MustPath("/opt", false, testtool),
+				), ignorePathname, wantChecksumOffline, nil},
+
+				{"error passthrough", pkg.NewExec(
+					"", nil, 0, true,
+					pkg.AbsWork,
+					[]string{"HAKUREI_TEST=1"},
+					check.MustAbs("/opt/bin/testtool"),
+					[]string{"testtool"},
+
+					pkg.MustPath("/proc/nonexistent", false, &stubArtifact{
+						kind:   pkg.KindTar,
+						params: []byte("doomed artifact"),
+						cure: func(t *pkg.TContext) error {
+							return stub.UniqueError(0xcafe)
+						},
+					}),
+				), nil, pkg.Checksum{}, &pkg.DependencyCureError{
+					{
+						Ident: unique.Make(pkg.ID(pkg.MustDecode(
+							"CWEoJqnSBpWf8uryC2qnIe3O1a_FZWUWZGbiVPsQFGW7pvDHiSwoK3QCU9-uxN87",
+						))),
+						Err: stub.UniqueError(0xcafe),
+					},
+				}},
+
+				{"invalid paths", pkg.NewExec(
+					"", nil, 0, false,
+					pkg.AbsWork,
+					[]string{"HAKUREI_TEST=1"},
+					check.MustAbs("/opt/bin/testtool"),
+					[]string{"testtool"},
+
+					pkg.ExecPath{},
+				), nil, pkg.Checksum{}, os.ErrInvalid},
+			})
+
+			// check init failure passthrough
+			var exitError *exec.ExitError
+			if _, _, err := c.Cure(pkg.NewExec(
+				"", nil, 0, false,
+				pkg.AbsWork,
+				nil,
+				check.MustAbs("/opt/bin/testtool"),
+				[]string{"testtool"},
+			)); !errors.As(err, &exitError) ||
+				exitError.ExitCode() != hst.ExitFailure {
+				t.Fatalf("Cure: error = %v, want init exit status 1", err)
+			}
+
+			testtoolDestroy(t, base, c)
+		}, pkg.MustDecode("UiV6kMz7KrTsc_yphiyQzFLqjRanHxUOwrBMtkKuWo4mOO6WgPFAcoUEeSp7eVIW")},
+
+		{"net", nil, func(t *testing.T, base *check.Absolute, c *pkg.Cache) {
+			c.SetStrict(true)
+			testtool, testtoolDestroy := newTesttool()
+
+			wantChecksum := pkg.MustDecode(
+				"a1F_i9PVQI4qMcoHgTQkORuyWLkC1GLIxOhDt2JpU1NGAxWc5VJzdlfRK-PYBh3W",
+			)
+			cureMany(t, c, []cureStep{
+				{"container", pkg.NewExec(
+					"exec-net", &wantChecksum, 0, false,
+					pkg.AbsWork,
+					[]string{"HAKUREI_TEST=1"},
+					check.MustAbs("/opt/bin/testtool"),
+					[]string{"testtool", "net"},
+
+					pkg.MustPath("/file", false, newStubFile(
+						pkg.KindHTTPGet,
+						pkg.ID{0xfe, 0},
+						nil,
+						nil, nil,
+					)),
+					pkg.MustPath("/.hakurei", false, &stubArtifact{
+						kind:   pkg.KindTar,
+						params: []byte("empty directory"),
+						cure: func(t *pkg.TContext) error {
+							return os.MkdirAll(t.GetWorkDir().String(), 0700)
+						},
+					}),
+					pkg.MustPath("/opt", false, testtool),
+				), ignorePathname, wantChecksum, nil},
+			})
+
+			testtoolDestroy(t, base, c)
+		}, pkg.MustDecode("ek4K-0d4iRSArkY2TCs3WK34DbiYeOmhE_4vsJTSu_6roY4ZF3YG6eKRooal-i1o")},
+
+		{"overlay root", nil, func(t *testing.T, base *check.Absolute, c *pkg.Cache) {
+			c.SetStrict(true)
+			testtool, testtoolDestroy := newTesttool()
+
+			cureMany(t, c, []cureStep{
+				{"container", pkg.NewExec(
+					"exec-overlay-root", nil, 0, false,
+					pkg.AbsWork,
+					[]string{"HAKUREI_TEST=1", "HAKUREI_ROOT=1"},
+					check.MustAbs("/opt/bin/testtool"),
+					[]string{"testtool"},
+
+					pkg.MustPath("/", true, &stubArtifact{
+						kind:   pkg.KindTar,
+						params: []byte("empty directory"),
+						cure: func(t *pkg.TContext) error {
+							return os.MkdirAll(t.GetWorkDir().String(), 0700)
+						},
+					}),
+					pkg.MustPath("/opt", false, testtool),
+				), ignorePathname, wantChecksumOffline, nil},
+			})
+
+			testtoolDestroy(t, base, c)
+		}, pkg.MustDecode("VIqqpf0ip9jcyw63i6E8lCMGUcLivQBe4Bevt3WusNac-1MSy5bzB647qGUBzl-W")},
+
+		{"overlay work", nil, func(t *testing.T, base *check.Absolute, c *pkg.Cache) {
+			c.SetStrict(true)
+			testtool, testtoolDestroy := newTesttool()
+
+			cureMany(t, c, []cureStep{
+				{"container", pkg.NewExec(
+					"exec-overlay-work", nil, 0, false,
+					pkg.AbsWork,
+					[]string{"HAKUREI_TEST=1", "HAKUREI_ROOT=1"},
+					check.MustAbs("/work/bin/testtool"),
+					[]string{"testtool"},
+
+					pkg.MustPath("/", true, &stubArtifact{
+						kind:   pkg.KindTar,
+						params: []byte("empty directory"),
+						cure: func(t *pkg.TContext) error {
+							return os.MkdirAll(t.GetWorkDir().String(), 0700)
+						},
+					}), pkg.MustPath("/work/", false, &stubArtifact{
+						kind:   pkg.KindTar,
+						params: []byte("empty directory"),
+						cure: func(t *pkg.TContext) error {
+							return os.MkdirAll(t.GetWorkDir().String(), 0700)
+						},
+					}), pkg.Path(pkg.AbsWork, false /* ignored */, testtool),
+				), ignorePathname, wantChecksumOffline, nil},
+			})
+
+			testtoolDestroy(t, base, c)
+		}, pkg.MustDecode("q8x2zQg4YZbKpPqKlEBj_uxXD9vOBaZ852qOuIsl9QdO73I_UMNpuUoPLtunxUYl")},
+
+		{"multiple layers", nil, func(t *testing.T, base *check.Absolute, c *pkg.Cache) {
+			c.SetStrict(true)
+			testtool, testtoolDestroy := newTesttool()
+
+			cureMany(t, c, []cureStep{
+				{"container", pkg.NewExec(
+					"exec-multiple-layers", nil, 0, false,
+					pkg.AbsWork,
+					[]string{"HAKUREI_TEST=1", "HAKUREI_ROOT=1"},
+					check.MustAbs("/opt/bin/testtool"),
+					[]string{"testtool", "layers"},
+
+					pkg.MustPath("/", true, &stubArtifact{
+						kind:   pkg.KindTar,
+						params: []byte("empty directory"),
+						cure: func(t *pkg.TContext) error {
+							return os.MkdirAll(t.GetWorkDir().String(), 0700)
+						},
+					}, &stubArtifactF{
+						kind:   pkg.KindExec,
+						params: []byte("test sample with dependencies"),
+
+						deps: slices.Repeat([]pkg.Artifact{newStubFile(
+							pkg.KindHTTPGet,
+							pkg.ID{0xfe, 0},
+							nil,
+							nil, nil,
+						), &stubArtifact{
+							kind:   pkg.KindTar,
+							params: []byte("empty directory"),
+
+							// this is queued and might run instead of the other
+							// one so do not leave it as nil
+							cure: func(t *pkg.TContext) error {
+								return os.MkdirAll(t.GetWorkDir().String(), 0700)
+							},
+						}}, 1<<5 /* concurrent cache hits */),
+
+						cure: func(f *pkg.FContext) error {
+							work := f.GetWorkDir()
+							if err := os.MkdirAll(work.String(), 0700); err != nil {
+								return err
+							}
+							return os.WriteFile(work.Append("check").String(), []byte("layers"), 0400)
+						},
+					}),
+					pkg.MustPath("/opt", false, testtool),
+				), ignorePathname, wantChecksumOffline, nil},
+			})
+
+			testtoolDestroy(t, base, c)
+		}, pkg.MustDecode("SITnQ6PTV12PAQQjIuLUxkvsXQiC9Gq_HJQlcb4BPL5YnRHnx8lsW7PRM9YMLBsx")},
+
+		{"overlay layer promotion", nil, func(t *testing.T, base *check.Absolute, c *pkg.Cache) {
+			c.SetStrict(true)
+			testtool, testtoolDestroy := newTesttool()
+
+			cureMany(t, c, []cureStep{
+				{"container", pkg.NewExec(
+					"exec-layer-promotion", nil, 0, true,
+					pkg.AbsWork,
+					[]string{"HAKUREI_TEST=1", "HAKUREI_ROOT=1"},
+					check.MustAbs("/opt/bin/testtool"),
+					[]string{"testtool", "promote"},
+
+					pkg.MustPath("/", true, &stubArtifact{
+						kind:   pkg.KindTar,
+						params: []byte("another empty directory"),
+						cure: func(t *pkg.TContext) error {
+							return os.MkdirAll(t.GetWorkDir().String(), 0700)
+						},
+					}, &stubArtifact{
+						kind:   pkg.KindTar,
+						params: []byte("empty directory"),
+						cure: func(t *pkg.TContext) error {
+							return os.MkdirAll(t.GetWorkDir().String(), 0700)
+						},
+					}),
+					pkg.MustPath("/opt", false, testtool),
+				), ignorePathname, wantChecksumOffline, nil},
+			})
+
+			testtoolDestroy(t, base, c)
+		}, pkg.MustDecode("fuC20BhMKr86TYzNPP2A-9P7mGLvdcOiG10exlhRvZm8ySI7csf0LhW3im_26l1N")},
+	})
+}
+
+// newTesttool returns an [Artifact] that cures into testtoolBin. The returned
+// function must be called at the end of the test but not deferred.
+func newTesttool() (
+	testtool pkg.Artifact,
+	testtoolDestroy func(t *testing.T, base *check.Absolute, c *pkg.Cache),
+) {
+	// testtoolBin is built during go:generate and is not deterministic
+	testtool = overrideIdent{pkg.ID{0xfe, 0xff}, &stubArtifact{
+		kind: pkg.KindTar,
+		cure: func(t *pkg.TContext) error {
+			work := t.GetWorkDir()
+			if err := os.MkdirAll(
+				work.Append("bin").String(),
+				0700,
+			); err != nil {
+				return err
+			}
+
+			if ift, err := net.Interfaces(); err != nil {
+				return err
+			} else {
+				var f *os.File
+				if f, err = os.Create(t.GetWorkDir().Append(
+					"ift",
+				).String()); err != nil {
+					return err
+				} else {
+					err = gob.NewEncoder(f).Encode(ift)
+					closeErr := f.Close()
+					if err != nil {
+						return err
+					}
+					if closeErr != nil {
+						return closeErr
+					}
+				}
+			}
+
+			return os.WriteFile(t.GetWorkDir().Append(
+				"bin",
+				"testtool",
+			).String(), testtoolBin, 0500)
+		},
+	}}
+	testtoolDestroy = newDestroyArtifactFunc(testtool)
+	return
+}

internal/pkg/file.go

@@ -0,0 +1,61 @@
+package pkg
+
+import (
+	"bytes"
+	"crypto/sha512"
+	"fmt"
+	"io"
+)
+
+// A fileArtifact is an [Artifact] that cures into data known ahead of time.
+type fileArtifact []byte
+
+var _ KnownChecksum = new(fileArtifact)
+
+// fileArtifactNamed embeds fileArtifact alongside a caller-supplied name.
+type fileArtifactNamed struct {
+	fileArtifact
+	// Caller-supplied user-facing reporting name.
+	name string
+}
+
+var _ fmt.Stringer = new(fileArtifactNamed)
+var _ KnownChecksum = new(fileArtifactNamed)
+
+// String returns the caller-supplied reporting name.
+func (a *fileArtifactNamed) String() string { return a.name }
+
+// NewFile returns a [FileArtifact] that cures into a caller-supplied byte slice.
+//
+// Caller must not modify data after NewFile returns.
+func NewFile(name string, data []byte) FileArtifact {
+	f := fileArtifact(data)
+	if name != "" {
+		return &fileArtifactNamed{f, name}
+	}
+	return &f
+}
+
+// Kind returns the hardcoded [Kind] constant.
+func (*fileArtifact) Kind() Kind { return KindFile }
+
+// Params writes the result of Cure.
+func (a *fileArtifact) Params(ctx *IContext) { ctx.GetHash().Write(*a) }
+
+// Dependencies returns a nil slice.
+func (*fileArtifact) Dependencies() []Artifact { return nil }
+
+// IsExclusive returns false: Cure returns a prepopulated buffer.
+func (*fileArtifact) IsExclusive() bool { return false }
+
+// Checksum computes and returns the checksum of caller-supplied data.
+func (a *fileArtifact) Checksum() Checksum {
+	h := sha512.New384()
+	h.Write(*a)
+	return Checksum(h.Sum(nil))
+}
+
+// Cure returns the caller-supplied data.
+func (a *fileArtifact) Cure(*RContext) (io.ReadCloser, error) {
+	return io.NopCloser(bytes.NewReader(*a)), nil
+}

internal/pkg/file_test.go

@@ -0,0 +1,29 @@
+package pkg_test
+
+import (
+	"testing"
+
+	"hakurei.app/container/check"
+	"hakurei.app/internal/pkg"
+)
+
+func TestFile(t *testing.T) {
+	t.Parallel()
+
+	checkWithCache(t, []cacheTestCase{
+		{"file", nil, func(t *testing.T, base *check.Absolute, c *pkg.Cache) {
+			c.SetStrict(true)
+
+			cureMany(t, c, []cureStep{
+				{"short", pkg.NewFile("null", []byte{0}), base.Append(
+					"identifier",
+					"lIx_W4M7tVOcQ8jh08EJOfXf4brRmkEEjvUa7c17vVUzlmtUxlhhrgqmc9aZhjbn",
+				), pkg.MustDecode(
+					"vsAhtPNo4waRNOASwrQwcIPTqb3SBuJOXw2G4T1mNmVZM-wrQTRllmgXqcIIoRcX",
+				), nil},
+			})
+		}, pkg.MustDecode(
+			"hnrfmJtivNKcgtETsKnU9gP_OwPgpNY3DSUJnmxnmeOODSO-YBvEBiTgieY4AAd7",
+		)},
+	})
+}

internal/pkg/net.go

@@ -0,0 +1,94 @@
+package pkg
+
+import (
+	"fmt"
+	"io"
+	"net/http"
+	"path"
+	"unique"
+)
+
+// An httpArtifact is an [Artifact] backed by a [http] url string. The method is
+// hardcoded as [http.MethodGet]. Request body is not allowed because it cannot
+// be deterministically represented by Params.
+type httpArtifact struct {
+	// Caller-supplied url string.
+	url string
+
+	// Caller-supplied checksum of the response body. This is validated when
+	// closing the [io.ReadCloser] returned by Cure.
+	checksum unique.Handle[Checksum]
+
+	// doFunc is the Do method of [http.Client] supplied by the caller.
+	doFunc func(req *http.Request) (*http.Response, error)
+}
+
+var _ KnownChecksum = new(httpArtifact)
+var _ fmt.Stringer = new(httpArtifact)
+
+// NewHTTPGet returns a new [FileArtifact] backed by the supplied client. A GET
+// request is set up for url. If c is nil, [http.DefaultClient] is used instead.
+func NewHTTPGet(
+	c *http.Client,
+	url string,
+	checksum Checksum,
+) FileArtifact {
+	if c == nil {
+		c = http.DefaultClient
+	}
+	return &httpArtifact{url: url, checksum: unique.Make(checksum), doFunc: c.Do}
+}
+
+// Kind returns the hardcoded [Kind] constant.
+func (*httpArtifact) Kind() Kind { return KindHTTPGet }
+
+// Params writes the backing url string. Client is not represented as it does
+// not affect [Cache.Cure] outcome.
+func (a *httpArtifact) Params(ctx *IContext) {
+	ctx.GetHash().Write([]byte(a.url))
+}
+
+// Dependencies returns a nil slice.
+func (*httpArtifact) Dependencies() []Artifact { return nil }
+
+// IsExclusive returns false: Cure returns as soon as a response is received.
+func (*httpArtifact) IsExclusive() bool { return false }
+
+// Checksum returns the caller-supplied checksum.
+func (a *httpArtifact) Checksum() Checksum { return a.checksum.Value() }
+
+// String returns [path.Base] over the backing url.
+func (a *httpArtifact) String() string { return path.Base(a.url) }
+
+// ResponseStatusError is returned for a response returned by an [http.Client]
+// with a status code other than [http.StatusOK].
+type ResponseStatusError int
+
+func (e ResponseStatusError) Error() string {
+	return "the requested URL returned non-OK status: " + http.StatusText(int(e))
+}
+
+// Cure sends the http request and returns the resulting response body reader
+// wrapped to perform checksum validation. It is valid but not encouraged to
+// close the resulting [io.ReadCloser] before it is read to EOF, as that causes
+// Close to block until all remaining data is consumed and validated.
+func (a *httpArtifact) Cure(r *RContext) (rc io.ReadCloser, err error) {
+	var req *http.Request
+	req, err = http.NewRequestWithContext(r.Unwrap(), http.MethodGet, a.url, nil)
+	if err != nil {
+		return
+	}
+
+	var resp *http.Response
+	if resp, err = a.doFunc(req); err != nil {
+		return
+	}
+
+	if resp.StatusCode != http.StatusOK {
+		_ = resp.Body.Close()
+		return nil, ResponseStatusError(resp.StatusCode)
+	}
+
+	rc = r.NewMeasuredReader(resp.Body, a.checksum)
+	return
+}

internal/pkg/net_test.go

@@ -0,0 +1,161 @@
+package pkg_test
+
+import (
+	"crypto/sha512"
+	"io"
+	"net/http"
+	"reflect"
+	"testing"
+	"testing/fstest"
+	"unique"
+	"unsafe"
+
+	"hakurei.app/container/check"
+	"hakurei.app/internal/pkg"
+)
+
+func TestHTTPGet(t *testing.T) {
+	t.Parallel()
+
+	const testdata = "\x7f\xe1\x69\xa2\xdd\x63\x96\x26\x83\x79\x61\x8b\xf0\x3f\xd5\x16\x9a\x39\x3a\xdb\xcf\xb1\xbc\x8d\x33\xff\x75\xee\x62\x56\xa9\xf0\x27\xac\x13\x94\x69"
+
+	testdataChecksum := func() unique.Handle[pkg.Checksum] {
+		h := sha512.New384()
+		h.Write([]byte(testdata))
+		return unique.Make(pkg.Checksum(h.Sum(nil)))
+	}()
+
+	var transport http.Transport
+	client := http.Client{Transport: &transport}
+	transport.RegisterProtocol("file", http.NewFileTransportFS(fstest.MapFS{
+		"testdata": {Data: []byte(testdata), Mode: 0400},
+	}))
+
+	checkWithCache(t, []cacheTestCase{
+		{"direct", nil, func(t *testing.T, base *check.Absolute, c *pkg.Cache) {
+			var r pkg.RContext
+			rCacheVal := reflect.ValueOf(&r).Elem().FieldByName("cache")
+			reflect.NewAt(
+				rCacheVal.Type(),
+				unsafe.Pointer(rCacheVal.UnsafeAddr()),
+			).Elem().Set(reflect.ValueOf(c))
+
+			f := pkg.NewHTTPGet(
+				&client,
+				"file:///testdata",
+				testdataChecksum.Value(),
+			)
+			var got []byte
+			if rc, err := f.Cure(&r); err != nil {
+				t.Fatalf("Cure: error = %v", err)
+			} else if got, err = io.ReadAll(rc); err != nil {
+				t.Fatalf("ReadAll: error = %v", err)
+			} else if string(got) != testdata {
+				t.Fatalf("Cure: %x, want %x", got, testdata)
+			} else if err = rc.Close(); err != nil {
+				t.Fatalf("Close: error = %v", err)
+			}
+
+			// check direct validation
+			f = pkg.NewHTTPGet(
+				&client,
+				"file:///testdata",
+				pkg.Checksum{},
+			)
+			wantErrMismatch := &pkg.ChecksumMismatchError{
+				Got: testdataChecksum.Value(),
+			}
+			if rc, err := f.Cure(&r); err != nil {
+				t.Fatalf("Cure: error = %v", err)
+			} else if got, err = io.ReadAll(rc); err != nil {
+				t.Fatalf("ReadAll: error = %v", err)
+			} else if string(got) != testdata {
+				t.Fatalf("Cure: %x, want %x", got, testdata)
+			} else if err = rc.Close(); !reflect.DeepEqual(err, wantErrMismatch) {
+				t.Fatalf("Close: error = %#v, want %#v", err, wantErrMismatch)
+			}
+
+			// check fallback validation
+			if rc, err := f.Cure(&r); err != nil {
+				t.Fatalf("Cure: error = %v", err)
+			} else if err = rc.Close(); !reflect.DeepEqual(err, wantErrMismatch) {
+				t.Fatalf("Close: error = %#v, want %#v", err, wantErrMismatch)
+			}
+
+			// check direct response error
+			f = pkg.NewHTTPGet(
+				&client,
+				"file:///nonexistent",
+				pkg.Checksum{},
+			)
+			wantErrNotFound := pkg.ResponseStatusError(http.StatusNotFound)
+			if _, err := f.Cure(&r); !reflect.DeepEqual(err, wantErrNotFound) {
+				t.Fatalf("Cure: error = %#v, want %#v", err, wantErrNotFound)
+			}
+		}, pkg.MustDecode("E4vEZKhCcL2gPZ2Tt59FS3lDng-d_2SKa2i5G_RbDfwGn6EemptFaGLPUDiOa94C")},
+
+		{"cure", nil, func(t *testing.T, base *check.Absolute, c *pkg.Cache) {
+			var r pkg.RContext
+			rCacheVal := reflect.ValueOf(&r).Elem().FieldByName("cache")
+			reflect.NewAt(
+				rCacheVal.Type(),
+				unsafe.Pointer(rCacheVal.UnsafeAddr()),
+			).Elem().Set(reflect.ValueOf(c))
+
+			f := pkg.NewHTTPGet(
+				&client,
+				"file:///testdata",
+				testdataChecksum.Value(),
+			)
+			wantPathname := base.Append(
+				"identifier",
+				"NqVORkT6L9HX6Za7kT2zcibY10qFqBaxEjPiYFrBQX-ZFr3yxCzJxbKOP0zVjeWb",
+			)
+			if pathname, checksum, err := c.Cure(f); err != nil {
+				t.Fatalf("Cure: error = %v", err)
+			} else if !pathname.Is(wantPathname) {
+				t.Fatalf("Cure: %q, want %q", pathname, wantPathname)
+			} else if checksum != testdataChecksum {
+				t.Fatalf("Cure: %x, want %x", checksum.Value(), testdataChecksum.Value())
+			}
+
+			var got []byte
+			if rc, err := f.Cure(&r); err != nil {
+				t.Fatalf("Cure: error = %v", err)
+			} else if got, err = io.ReadAll(rc); err != nil {
+				t.Fatalf("ReadAll: error = %v", err)
+			} else if string(got) != testdata {
+				t.Fatalf("Cure: %x, want %x", got, testdata)
+			} else if err = rc.Close(); err != nil {
+				t.Fatalf("Close: error = %v", err)
+			}
+
+			// check load from cache
+			f = pkg.NewHTTPGet(
+				&client,
+				"file:///testdata",
+				testdataChecksum.Value(),
+			)
+			if rc, err := f.Cure(&r); err != nil {
+				t.Fatalf("Cure: error = %v", err)
+			} else if got, err = io.ReadAll(rc); err != nil {
+				t.Fatalf("ReadAll: error = %v", err)
+			} else if string(got) != testdata {
+				t.Fatalf("Cure: %x, want %x", got, testdata)
+			} else if err = rc.Close(); err != nil {
+				t.Fatalf("Close: error = %v", err)
+			}
+
+			// check error passthrough
+			f = pkg.NewHTTPGet(
+				&client,
+				"file:///nonexistent",
+				pkg.Checksum{},
+			)
+			wantErrNotFound := pkg.ResponseStatusError(http.StatusNotFound)
+			if _, _, err := c.Cure(f); !reflect.DeepEqual(err, wantErrNotFound) {
+				t.Fatalf("Pathname: error = %#v, want %#v", err, wantErrNotFound)
+			}
+		}, pkg.MustDecode("bqtn69RkV5E7V7GhhgCFjcvbxmaqrO8DywamM4Tyjf10F6EJBHjXiIa_tFRtF4iN")},
+	})
+}

internal/pkg/tar.go

@@ -0,0 +1,243 @@
+package pkg
+
+import (
+	"archive/tar"
+	"compress/bzip2"
+	"compress/gzip"
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+	"io/fs"
+	"net/http"
+	"os"
+
+	"hakurei.app/container/check"
+)
+
+const (
+	// TarUncompressed denotes an uncompressed tarball.
+	TarUncompressed = iota
+	// TarGzip denotes a tarball compressed via [gzip].
+	TarGzip
+	// TarBzip2 denotes a tarball compressed via [bzip2].
+	TarBzip2
+)
+
+// A tarArtifact is an [Artifact] unpacking a tarball backed by a [FileArtifact].
+type tarArtifact struct {
+	// Caller-supplied backing tarball.
+	f Artifact
+	// Compression on top of the tarball.
+	compression uint64
+}
+
+// tarArtifactNamed embeds tarArtifact for a [fmt.Stringer] tarball.
+type tarArtifactNamed struct {
+	tarArtifact
+	// Copied from tarArtifact.f.
+	name string
+}
+
+var _ fmt.Stringer = new(tarArtifactNamed)
+
+// String returns the name of the underlying [Artifact] suffixed with unpack.
+func (a *tarArtifactNamed) String() string { return a.name + "-unpack" }
+
+// NewTar returns a new [Artifact] backed by the supplied [Artifact] and
+// compression method. The source [Artifact] must be compatible with
+// [TContext.Open].
+func NewTar(a Artifact, compression uint64) Artifact {
+	ta := tarArtifact{a, compression}
+	if s, ok := a.(fmt.Stringer); ok {
+		if name := s.String(); name != "" {
+			return &tarArtifactNamed{ta, name}
+		}
+	}
+	return &ta
+}
+
+// NewHTTPGetTar is abbreviation for NewHTTPGet passed to NewTar.
+func NewHTTPGetTar(
+	hc *http.Client,
+	url string,
+	checksum Checksum,
+	compression uint64,
+) Artifact {
+	return NewTar(NewHTTPGet(hc, url, checksum), compression)
+}
+
+// Kind returns the hardcoded [Kind] constant.
+func (a *tarArtifact) Kind() Kind { return KindTar }
+
+// Params writes compression encoded in little endian.
+func (a *tarArtifact) Params(ctx *IContext) {
+	ctx.GetHash().Write(binary.LittleEndian.AppendUint64(nil, a.compression))
+}
+
+// Dependencies returns a slice containing the backing file.
+func (a *tarArtifact) Dependencies() []Artifact {
+	return []Artifact{a.f}
+}
+
+// IsExclusive returns false: decompressor and tar reader are fully sequential.
+func (a *tarArtifact) IsExclusive() bool { return false }
+
+// A DisallowedTypeflagError describes a disallowed typeflag encountered while
+// unpacking a tarball.
+type DisallowedTypeflagError byte
+
+func (e DisallowedTypeflagError) Error() string {
+	return "disallowed typeflag '" + string(e) + "'"
+}
+
+// Cure cures the [Artifact], producing a directory located at work.
+func (a *tarArtifact) Cure(t *TContext) (err error) {
+	temp := t.GetTempDir()
+	var tr io.ReadCloser
+	if tr, err = t.Open(a.f); err != nil {
+		return
+	}
+
+	defer func(f io.ReadCloser) {
+		if err == nil {
+			err = tr.Close()
+		}
+
+		closeErr := f.Close()
+		if err == nil {
+			err = closeErr
+		}
+	}(tr)
+	tr = io.NopCloser(tr)
+
+	switch a.compression {
+	case TarUncompressed:
+		break
+
+	case TarGzip:
+		if tr, err = gzip.NewReader(tr); err != nil {
+			return
+		}
+		break
+
+	case TarBzip2:
+		tr = io.NopCloser(bzip2.NewReader(tr))
+		break
+
+	default:
+		return os.ErrInvalid
+	}
+
+	type dirTargetPerm struct {
+		path *check.Absolute
+		mode fs.FileMode
+	}
+	var madeDirectories []dirTargetPerm
+
+	if err = os.MkdirAll(temp.String(), 0700); err != nil {
+		return
+	}
+
+	var header *tar.Header
+	r := tar.NewReader(tr)
+	for header, err = r.Next(); err == nil; header, err = r.Next() {
+		typeflag := header.Typeflag
+		if typeflag == 0 {
+			if len(header.Name) > 0 && header.Name[len(header.Name)-1] == '/' {
+				typeflag = tar.TypeDir
+			} else {
+				typeflag = tar.TypeReg
+			}
+		}
+
+		pathname := temp.Append(header.Name)
+		if typeflag >= '0' && typeflag <= '9' && typeflag != tar.TypeDir {
+			if err = os.MkdirAll(pathname.Dir().String(), 0700); err != nil {
+				return
+			}
+		}
+
+		switch typeflag {
+		case tar.TypeReg:
+			var f *os.File
+			if f, err = os.OpenFile(
+				pathname.String(),
+				os.O_CREATE|os.O_EXCL|os.O_WRONLY,
+				header.FileInfo().Mode()&0500,
+			); err != nil {
+				return
+			}
+			if _, err = io.Copy(f, r); err != nil {
+				_ = f.Close()
+				return
+			} else if err = f.Close(); err != nil {
+				return
+			}
+			break
+
+		case tar.TypeLink:
+			if err = os.Link(
+				temp.Append(header.Linkname).String(),
+				pathname.String(),
+			); err != nil {
+				return
+			}
+			break
+
+		case tar.TypeSymlink:
+			if err = os.Symlink(header.Linkname, pathname.String()); err != nil {
+				return
+			}
+			break
+
+		case tar.TypeDir:
+			madeDirectories = append(madeDirectories, dirTargetPerm{
+				path: pathname,
+				mode: header.FileInfo().Mode(),
+			})
+			if err = os.MkdirAll(pathname.String(), 0700); err != nil {
+				return
+			}
+			break
+
+		case tar.TypeXGlobalHeader:
+			continue // ignore
+
+		default:
+			return DisallowedTypeflagError(typeflag)
+		}
+	}
+	if errors.Is(err, io.EOF) {
+		err = nil
+	}
+	if err == nil {
+		for _, e := range madeDirectories {
+			if err = os.Chmod(e.path.String(), e.mode&0500); err != nil {
+				return
+			}
+		}
+	} else {
+		return
+	}
+
+	if err = os.Chmod(temp.String(), 0700); err != nil {
+		return
+	}
+
+	var entries []os.DirEntry
+	if entries, err = os.ReadDir(temp.String()); err != nil {
+		return
+	}
+
+	if len(entries) == 1 && entries[0].IsDir() {
+		p := temp.Append(entries[0].Name())
+		if err = os.Chmod(p.String(), 0700); err != nil {
+			return
+		}
+		err = os.Rename(p.String(), t.GetWorkDir().String())
+	} else {
+		err = os.Rename(temp.String(), t.GetWorkDir().String())
+	}
+	return
+}

internal/pkg/tar_test.go

@@ -0,0 +1,203 @@
+package pkg_test
+
+import (
+	"archive/tar"
+	"bytes"
+	"compress/gzip"
+	"crypto/sha512"
+	"errors"
+	"io/fs"
+	"net/http"
+	"os"
+	"testing"
+	"testing/fstest"
+
+	"hakurei.app/container/check"
+	"hakurei.app/container/stub"
+	"hakurei.app/internal/pkg"
+)
+
+func TestTar(t *testing.T) {
+	t.Parallel()
+
+	checkWithCache(t, []cacheTestCase{
+		{"http", nil, func(t *testing.T, base *check.Absolute, c *pkg.Cache) {
+			checkTarHTTP(t, base, c, fstest.MapFS{
+				".": {Mode: fs.ModeDir | 0700},
+
+				"checksum": {Mode: fs.ModeDir | 0700},
+				"checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP":                {Mode: fs.ModeDir | 0700},
+				"checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP/check":          {Mode: 0400, Data: []byte{0, 0}},
+				"checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP/lib":            {Mode: fs.ModeDir | 0700},
+				"checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP/lib/pkgconfig":  {Mode: fs.ModeDir | 0700},
+				"checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP/lib/libedac.so": {Mode: fs.ModeSymlink | 0777, Data: []byte("/proc/nonexistent/libedac.so")},
+
+				"identifier": {Mode: fs.ModeDir | 0700},
+				"identifier/HnySzeLQvSBZuTUcvfmLEX_OmH4yJWWH788NxuLuv7kVn8_uPM6Ks4rqFWM2NZJY": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP")},
+				"identifier/Zx5ZG9BAwegNT3zQwCySuI2ktCXxNgxirkGLFjW4FW06PtojYVaCdtEw8yuntPLa": {Mode: fs.ModeSymlink | 0777, Data: []byte("../checksum/1TL00Qb8dcqayX7wTO8WNaraHvY6b-KCsctLDTrb64QBCmxj_-byK1HdIUwMaFEP")},
+
+				"work": {Mode: fs.ModeDir | 0700},
+			}, pkg.MustDecode(
+				"cTw0h3AmYe7XudSoyEMByduYXqGi-N5ZkTZ0t9K5elsu3i_jNIVF5T08KR1roBFM",
+			))
+		}, pkg.MustDecode("sxbgyX-bPoezbha214n2lbQhiVfTUBkhZ0EX6zI7mmkMdrCdwuMwhMBJphLQsy94")},
+
+		{"http expand", nil, func(t *testing.T, base *check.Absolute, c *pkg.Cache) {
+			checkTarHTTP(t, base, c, fstest.MapFS{
+				".": {Mode: fs.ModeDir | 0700},
+
+				"lib":            {Mode: fs.ModeDir | 0700},
+				"lib/libedac.so": {Mode: fs.ModeSymlink | 0777, Data: []byte("/proc/nonexistent/libedac.so")},
+			}, pkg.MustDecode(
+				"CH3AiUrCCcVOjOYLaMKKK1Da78989JtfHeIsxMzWOQFiN4mrCLDYpoDxLWqJWCUN",
+			))
+		}, pkg.MustDecode("4I8wx_h7NSJTlG5lbuz-GGEXrOg0GYC3M_503LYEBhv5XGWXfNIdIY9Q3eVSYldX")},
+	})
+}
+
+func checkTarHTTP(
+	t *testing.T,
+	base *check.Absolute,
+	c *pkg.Cache,
+	testdataFsys fs.FS,
+	wantChecksum pkg.Checksum,
+) {
+	var testdata string
+	{
+		var buf bytes.Buffer
+		w := tar.NewWriter(&buf)
+		if err := w.AddFS(testdataFsys); err != nil {
+			t.Fatalf("AddFS: error = %v", err)
+		}
+		if err := w.Close(); err != nil {
+			t.Fatalf("Close: error = %v", err)
+		}
+
+		var zbuf bytes.Buffer
+		gw := gzip.NewWriter(&zbuf)
+		if _, err := gw.Write(buf.Bytes()); err != nil {
+			t.Fatalf("Write: error = %v", err)
+		}
+		if err := gw.Close(); err != nil {
+			t.Fatalf("Close: error = %v", err)
+		}
+		testdata = zbuf.String()
+	}
+
+	testdataChecksum := func() pkg.Checksum {
+		h := sha512.New384()
+		h.Write([]byte(testdata))
+		return (pkg.Checksum)(h.Sum(nil))
+	}()
+
+	var transport http.Transport
+	client := http.Client{Transport: &transport}
+	transport.RegisterProtocol("file", http.NewFileTransportFS(fstest.MapFS{
+		"testdata": {Data: []byte(testdata), Mode: 0400},
+	}))
+
+	wantIdent := func() pkg.ID {
+		h := sha512.New384()
+		h.Write([]byte{byte(pkg.KindTar), 0, 0, 0, 0, 0, 0, 0})
+		h.Write([]byte{pkg.TarGzip, 0, 0, 0, 0, 0, 0, 0})
+		h.Write([]byte{byte(pkg.KindHTTPGet), 0, 0, 0, 0, 0, 0, 0})
+
+		h0 := sha512.New384()
+		h0.Write([]byte{byte(pkg.KindHTTPGet), 0, 0, 0, 0, 0, 0, 0})
+		h0.Write([]byte("file:///testdata"))
+		h.Write(h0.Sum(nil))
+		return pkg.ID(h.Sum(nil))
+	}()
+
+	a := pkg.NewHTTPGetTar(
+		&client,
+		"file:///testdata",
+		testdataChecksum,
+		pkg.TarGzip,
+	)
+
+	tarDir := stubArtifact{
+		kind:   pkg.KindExec,
+		params: []byte("directory containing a single regular file"),
+		cure: func(t *pkg.TContext) error {
+			work := t.GetWorkDir()
+			if err := os.MkdirAll(work.String(), 0700); err != nil {
+				return err
+			}
+			return os.WriteFile(
+				work.Append("sample.tar.gz").String(),
+				[]byte(testdata),
+				0400,
+			)
+		},
+	}
+	tarDirMulti := stubArtifact{
+		kind:   pkg.KindExec,
+		params: []byte("directory containing a multiple entries"),
+		cure: func(t *pkg.TContext) error {
+			work := t.GetWorkDir()
+			if err := os.MkdirAll(work.Append(
+				"garbage",
+			).String(), 0700); err != nil {
+				return err
+			}
+			return os.WriteFile(
+				work.Append("sample.tar.gz").String(),
+				[]byte(testdata),
+				0400,
+			)
+		},
+	}
+	tarDirType := stubArtifact{
+		kind:   pkg.KindExec,
+		params: []byte("directory containing a symbolic link"),
+		cure: func(t *pkg.TContext) error {
+			work := t.GetWorkDir()
+			if err := os.MkdirAll(work.String(), 0700); err != nil {
+				return err
+			}
+			return os.Symlink(
+				work.String(),
+				work.Append("sample.tar.gz").String(),
+			)
+		},
+	}
+	// destroy these to avoid including it in flatten test case
+	defer newDestroyArtifactFunc(&tarDir)(t, base, c)
+	defer newDestroyArtifactFunc(&tarDirMulti)(t, base, c)
+	defer newDestroyArtifactFunc(&tarDirType)(t, base, c)
+
+	cureMany(t, c, []cureStep{
+		{"file", a, base.Append(
+			"identifier",
+			pkg.Encode(wantIdent),
+		), wantChecksum, nil},
+
+		{"directory", pkg.NewTar(
+			&tarDir,
+			pkg.TarGzip,
+		), ignorePathname, wantChecksum, nil},
+
+		{"multiple entries", pkg.NewTar(
+			&tarDirMulti,
+			pkg.TarGzip,
+		), nil, pkg.Checksum{}, errors.New(
+			"input directory does not contain a single regular file",
+		)},
+
+		{"bad type", pkg.NewTar(
+			&tarDirType,
+			pkg.TarGzip,
+		), nil, pkg.Checksum{}, errors.New(
+			"input directory does not contain a single regular file",
+		)},
+
+		{"error passthrough", pkg.NewTar(&stubArtifact{
+			kind:   pkg.KindExec,
+			params: []byte("doomed artifact"),
+			cure: func(t *pkg.TContext) error {
+				return stub.UniqueError(0xcafe)
+			},
+		}, pkg.TarGzip), nil, pkg.Checksum{}, stub.UniqueError(0xcafe)},
+	})
+}

internal/pkg/testdata/main.go

@@ -0,0 +1,268 @@
+//go:build testtool
+
+package main
+
+import (
+	"encoding/gob"
+	"log"
+	"net"
+	"os"
+	"path"
+	"reflect"
+	"slices"
+	"strings"
+
+	"hakurei.app/container/check"
+	"hakurei.app/container/fhs"
+	"hakurei.app/container/vfs"
+)
+
+func main() {
+	log.SetFlags(0)
+	log.SetPrefix("testtool: ")
+
+	var hostNet, layers, promote bool
+	if len(os.Args) == 2 && os.Args[0] == "testtool" {
+		switch os.Args[1] {
+		case "net":
+			hostNet = true
+			log.SetPrefix("testtool(net): ")
+			break
+
+		case "layers":
+			layers = true
+			log.SetPrefix("testtool(layers): ")
+			break
+
+		case "promote":
+			promote = true
+			log.SetPrefix("testtool(promote): ")
+
+		default:
+			log.Fatalf("Args: %q", os.Args)
+			return
+		}
+	} else if wantArgs := []string{"testtool"}; !slices.Equal(os.Args, wantArgs) {
+		log.Fatalf("Args: %q, want %q", os.Args, wantArgs)
+	}
+
+	var overlayRoot bool
+	wantEnv := []string{"HAKUREI_TEST=1"}
+	if len(os.Environ()) == 2 {
+		overlayRoot = true
+		if !layers && !promote {
+			log.SetPrefix("testtool(overlay root): ")
+		}
+		wantEnv = []string{"HAKUREI_TEST=1", "HAKUREI_ROOT=1"}
+	}
+	if !slices.Equal(wantEnv, os.Environ()) {
+		log.Fatalf("Environ: %q, want %q", os.Environ(), wantEnv)
+	}
+
+	var overlayWork bool
+	const (
+		wantExec     = "/opt/bin/testtool"
+		wantExecWork = "/work/bin/testtool"
+	)
+	var iftPath string
+	if got, err := os.Executable(); err != nil {
+		log.Fatalf("Executable: error = %v", err)
+	} else {
+		iftPath = path.Join(path.Dir(path.Dir(got)), "ift")
+
+		if got != wantExec {
+			switch got {
+			case wantExecWork:
+				overlayWork = true
+				log.SetPrefix("testtool(overlay work): ")
+
+			default:
+				log.Fatalf("Executable: %q, want %q", got, wantExec)
+			}
+		}
+	}
+
+	wantHostname := "cure"
+	if hostNet {
+		wantHostname += "-net"
+	}
+
+	if hostname, err := os.Hostname(); err != nil {
+		log.Fatalf("Hostname: error = %v", err)
+	} else if hostname != wantHostname {
+		log.Fatalf("Hostname: %q, want %q", hostname, wantHostname)
+	}
+
+	var m *vfs.MountInfo
+	if f, err := os.Open(fhs.Proc + "self/mountinfo"); err != nil {
+		log.Fatalf("Open: error = %v", err)
+	} else {
+		err = vfs.NewMountInfoDecoder(f).Decode(&m)
+		closeErr := f.Close()
+		if err != nil {
+			log.Fatalf("Decode: error = %v", err)
+		}
+		if closeErr != nil {
+			log.Fatalf("Close: error = %v", err)
+		}
+	}
+
+	if ift, err := net.Interfaces(); err != nil {
+		log.Fatal(err)
+	} else if !hostNet {
+		if len(ift) != 1 || ift[0].Name != "lo" {
+			log.Fatalln("got interfaces", strings.Join(slices.Collect(func(yield func(ifn string) bool) {
+				for _, ifi := range ift {
+					if !yield(ifi.Name) {
+						break
+					}
+				}
+			}), ", "))
+		}
+	} else {
+		var iftParent []net.Interface
+
+		var r *os.File
+		if r, err = os.Open(iftPath); err != nil {
+			log.Fatal(err)
+		} else {
+			err = gob.NewDecoder(r).Decode(&iftParent)
+			closeErr := r.Close()
+			if err != nil {
+				log.Fatal(err)
+			}
+			if closeErr != nil {
+				log.Fatal(closeErr)
+			}
+		}
+
+		if !reflect.DeepEqual(ift, iftParent) {
+			log.Fatalf("Interfaces: %#v, want %#v", ift, iftParent)
+		}
+	}
+
+	const checksumEmptyDir = "MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU"
+	ident := "U2cbgVgEtjfRuvHfE1cQnZ3t8yoexULQyo_VLgvxAVJSsobMcNaFIsuDWtmt7kzK"
+	log.Println(m)
+	next := func() { m = m.Next; log.Println(m) }
+
+	if overlayRoot {
+		ident = "5ey2wpmMpj483YYa7ZZQciYLA2cx3_l167JCqWW4Pd-5DVp81dj9EsBtVTwYptF6"
+
+		if m.Root != "/" || m.Target != "/" ||
+			m.Source != "overlay" || m.FsType != "overlay" {
+			log.Fatal("unexpected root mount entry")
+		}
+		var lowerdir string
+		for _, o := range strings.Split(m.FsOptstr, ",") {
+			const lowerdirKey = "lowerdir="
+			if strings.HasPrefix(o, lowerdirKey) {
+				lowerdir = o[len(lowerdirKey):]
+			}
+		}
+		if !layers {
+			if path.Base(lowerdir) != checksumEmptyDir {
+				log.Fatal("unexpected artifact checksum")
+			}
+		} else {
+			ident = "tfjrsVuBuFgzWgwz-yPppFtylYuC1VFWnKhyBiHbWTGkyz8lt7Ee9QXWaIHPXs4x"
+
+			lowerdirsEscaped := strings.Split(lowerdir, ":")
+			lowerdirs := lowerdirsEscaped[:0]
+			// ignore the option separator since it does not appear in ident
+			for i, e := range lowerdirsEscaped {
+				if len(e) > 0 &&
+					e[len(e)-1] == check.SpecialOverlayEscape[0] &&
+					(len(e) == 1 || e[len(e)-2] != check.SpecialOverlayEscape[0]) {
+					// ignore escaped pathname separator since it does not
+					// appear in ident
+
+					e = e[:len(e)-1]
+					if len(lowerdirsEscaped) != i {
+						lowerdirsEscaped[i+1] = e + lowerdirsEscaped[i+1]
+						continue
+					}
+				}
+				lowerdirs = append(lowerdirs, e)
+			}
+
+			if len(lowerdirs) != 2 ||
+				path.Base(lowerdirs[0]) != "MGWmEfjut2QE2xPJwTsmUzpff4BN_FEnQ7T0j7gvUCCiugJQNwqt9m151fm9D1yU" ||
+				path.Base(lowerdirs[1]) != "nY_CUdiaUM1OL4cPr5TS92FCJ3rCRV7Hm5oVTzAvMXwC03_QnTRfQ5PPs7mOU9fK" {
+				log.Fatalf("unexpected lowerdirs %s", strings.Join(lowerdirs, ", "))
+			}
+		}
+	} else {
+		if hostNet {
+			ident = "QdsJhGgnk5N2xdUNGcndXQxFKifxf1V_2t9X8CQ-pDcg24x6mGJC_BiLfGbs6Qml"
+		}
+
+		if m.Root != "/sysroot" || m.Target != "/" {
+			log.Fatal("unexpected root mount entry")
+		}
+
+		next()
+		if path.Base(m.Root) != "OLBgp1GsljhM2TJ-sbHjaiH9txEUvgdDTAzHv2P24donTt6_529l-9Ua0vFImLlb" {
+			log.Fatal("unexpected file artifact checksum")
+		}
+
+		next()
+		if path.Base(m.Root) != checksumEmptyDir {
+			log.Fatal("unexpected artifact checksum")
+		}
+	}
+
+	if promote {
+		ident = "O-6VjlIUxc4PYLf5v35uhIeL8kkYCbHYklqlmDjFPXe0m4j6GkUDg5qwTzBRESnf"
+	}
+
+	next() // testtool artifact
+
+	next()
+	if overlayWork {
+		ident = "acaDzHZv40dZaz4cGAXayqbRMgbEOuiuiUijZL8IgDQvyeCNMFE3onBMYfny-kXA"
+		if m.Root != "/" || m.Target != "/work" ||
+			m.Source != "overlay" || m.FsType != "overlay" {
+			log.Fatal("unexpected work mount entry")
+		}
+	} else {
+		if path.Base(m.Root) != ident || m.Target != "/work" {
+			log.Fatal("unexpected work mount entry")
+		}
+	}
+
+	next()
+	if path.Base(m.Root) != ident || m.Target != "/tmp" {
+		log.Fatal("unexpected temp mount entry")
+	}
+
+	next()
+	if m.Root != "/" || m.Target != "/proc" || m.Source != "proc" || m.FsType != "proc" {
+		log.Fatal("unexpected proc mount entry")
+	}
+
+	next()
+	if m.Root != "/" || m.Target != "/dev" || m.Source != "devtmpfs" || m.FsType != "tmpfs" {
+		log.Fatal("unexpected dev mount entry")
+	}
+
+	for i := 0; i < 9; i++ { // private /dev entries
+		next()
+	}
+
+	if m.Next != nil {
+		log.Println("unexpected extra mount entries")
+		for m.Next != nil {
+			next()
+		}
+		os.Exit(1)
+	}
+
+	checkData := []byte{0}
+	if hostNet {
+		checkData = []byte("net")
+	}
+	if err := os.WriteFile("check", checkData, 0400); err != nil {
+		log.Fatal(err)
+	}
+}

internal/rosa/acl.go

@@ -0,0 +1,98 @@
+package rosa
+
+import "hakurei.app/internal/pkg"
+
+func (t Toolchain) newAttr() pkg.Artifact {
+	const (
+		version  = "2.5.2"
+		checksum = "YWEphrz6vg1sUMmHHVr1CRo53pFXRhq_pjN-AlG8UgwZK1y6m7zuDhxqJhD0SV0l"
+	)
+	return t.New("attr-"+version, false, []pkg.Artifact{
+		t.Load(Make),
+		t.Load(Perl),
+	}, nil, nil, `
+ln -s ../../system/bin/perl /usr/bin
+
+cd "$(mktemp -d)"
+/usr/src/attr/configure \
+	--prefix=/system \
+	--build="${ROSA_TRIPLE}" \
+	--enable-static
+make "-j$(nproc)" check
+make DESTDIR=/work install
+`, pkg.Path(AbsUsrSrc.Append("attr"), true, t.NewPatchedSource(
+		"attr", version, pkg.NewHTTPGetTar(
+			nil,
+			"https://download.savannah.nongnu.org/releases/attr/"+
+				"attr-"+version+".tar.gz",
+			mustDecode(checksum),
+			pkg.TarGzip,
+		), true, [2]string{"libgen-basename", `From 8a80d895dfd779373363c3a4b62ecce5a549efb2 Mon Sep 17 00:00:00 2001
+From: "Haelwenn (lanodan) Monnier" <contact@hacktivis.me>
+Date: Sat, 30 Mar 2024 10:17:10 +0100
+Subject: tools/attr.c: Add missing libgen.h include for basename(3)
+
+Fixes compilation issue with musl and modern C99 compilers.
+
+See: https://bugs.gentoo.org/926294
+---
+ tools/attr.c | 1 +
+ 1 file changed, 1 insertion(+)
+
+diff --git a/tools/attr.c b/tools/attr.c
+index f12e4af..6a3c1e9 100644
+--- a/tools/attr.c
++++ b/tools/attr.c
+@@ -28,6 +28,7 @@
+ #include <errno.h>
+ #include <string.h>
+ #include <locale.h>
++#include <libgen.h>
+ 
+ #include <attr/attributes.h>
+ 
+-- 
+cgit v1.1`}, [2]string{"musl-errno", `diff --git a/test/attr.test b/test/attr.test
+index 6ce2f9b..e9bde92 100644
+--- a/test/attr.test
++++ b/test/attr.test
+@@ -11,7 +11,7 @@ Try various valid and invalid names
+
+        $ touch f
+        $ setfattr -n user -v value f
+-       > setfattr: f: Operation not supported
++       > setfattr: f: Not supported
+
+        $ setfattr -n user. -v value f
+        > setfattr: f: Invalid argument
+`},
+	)))
+}
+func init() { artifactsF[Attr] = Toolchain.newAttr }
+
+func (t Toolchain) newACL() pkg.Artifact {
+	const (
+		version  = "2.3.2"
+		checksum = "-fY5nwH4K8ZHBCRXrzLdguPkqjKI6WIiGu4dBtrZ1o0t6AIU73w8wwJz_UyjIS0P"
+	)
+	return t.New("acl-"+version, false, []pkg.Artifact{
+		t.Load(Make),
+
+		t.Load(Attr),
+	}, nil, nil, `
+cd "$(mktemp -d)"
+/usr/src/acl/configure \
+	--prefix=/system \
+	--build="${ROSA_TRIPLE}" \
+	--enable-static
+make "-j$(nproc)"
+make DESTDIR=/work install
+`, pkg.Path(AbsUsrSrc.Append("acl"), true, pkg.NewHTTPGetTar(
+		nil,
+		"https://download.savannah.nongnu.org/releases/acl/"+
+			"acl-"+version+".tar.gz",
+		mustDecode(checksum),
+		pkg.TarGzip,
+	)))
+}
+func init() { artifactsF[ACL] = Toolchain.newACL }

internal/rosa/all.go

@@ -0,0 +1,70 @@
+package rosa
+
+import (
+	"sync"
+
+	"hakurei.app/internal/pkg"
+)
+
+// PArtifact is a lazily-initialised [pkg.Artifact] preset.
+type PArtifact int
+
+const (
+	ACL PArtifact = iota
+	Attr
+	Autoconf
+	Bash
+	Busybox
+	CMake
+	Coreutils
+	Diffutils
+	Gettext
+	Git
+	Go
+	Gperf
+	Hakurei
+	KernelHeaders
+	LibXau
+	Libexpat
+	Libffi
+	Libgd
+	Libseccomp
+	Libxml2
+	M4
+	Make
+	Meson
+	Ninja
+	Patch
+	Perl
+	PkgConfig
+	Python
+	Rsync
+	Setuptools
+	Wayland
+	WaylandProtocols
+	XCB
+	XCBProto
+	Xproto
+	Zlib
+
+	// _presetEnd is the total number of presets and does not denote a preset.
+	_presetEnd
+)
+
+var (
+	// artifactsF is an array of functions for the result of [PArtifact].
+	artifactsF [_presetEnd]func(t Toolchain) pkg.Artifact
+
+	// artifacts stores the result of artifactsF.
+	artifacts [_toolchainEnd][len(artifactsF)]pkg.Artifact
+	// artifactsOnce is for lazy initialisation of artifacts.
+	artifactsOnce [_toolchainEnd][len(artifactsF)]sync.Once
+)
+
+// Load returns the resulting [pkg.Artifact] of [PArtifact].
+func (t Toolchain) Load(p PArtifact) pkg.Artifact {
+	artifactsOnce[t][p].Do(func() {
+		artifacts[t][p] = artifactsF[p](t)
+	})
+	return artifacts[t][p]
+}

internal/rosa/busybox.go

@@ -0,0 +1,359 @@
+package rosa
+
+import (
+	"fmt"
+	"io"
+	"net/http"
+	"os"
+	"slices"
+	"strings"
+	"time"
+
+	"hakurei.app/container/fhs"
+	"hakurei.app/internal/pkg"
+)
+
+// busyboxBin is a busybox binary distribution installed under bin/busybox.
+type busyboxBin struct {
+	// Underlying busybox binary.
+	bin pkg.FileArtifact
+}
+
+// Kind returns the hardcoded [pkg.Kind] value.
+func (a busyboxBin) Kind() pkg.Kind { return kindBusyboxBin }
+
+// Params is a noop.
+func (a busyboxBin) Params(*pkg.IContext) {}
+
+// IsExclusive returns false: Cure performs a trivial filesystem write.
+func (busyboxBin) IsExclusive() bool { return false }
+
+// Dependencies returns the underlying busybox [pkg.File].
+func (a busyboxBin) Dependencies() []pkg.Artifact {
+	return []pkg.Artifact{a.bin}
+}
+
+// String returns the reporting name of the underlying file prefixed with expand.
+func (a busyboxBin) String() string {
+	return "expand-" + a.bin.(fmt.Stringer).String()
+}
+
+// Cure installs the underlying busybox [pkg.File] to bin/busybox.
+func (a busyboxBin) Cure(t *pkg.TContext) (err error) {
+	var r io.ReadCloser
+	if r, err = t.Open(a.bin); err != nil {
+		return
+	}
+	defer func() {
+		closeErr := r.Close()
+		if err == nil {
+			err = closeErr
+		}
+	}()
+
+	binDir := t.GetWorkDir().Append("bin")
+	if err = os.MkdirAll(binDir.String(), 0700); err != nil {
+		return
+	}
+
+	var w *os.File
+	if w, err = os.OpenFile(
+		binDir.Append("busybox").String(),
+		os.O_WRONLY|os.O_CREATE|os.O_EXCL,
+		0500,
+	); err != nil {
+		return
+	}
+	defer func() {
+		closeErr := w.Close()
+		if err == nil {
+			err = closeErr
+		}
+	}()
+
+	_, err = io.Copy(w, r)
+	return
+}
+
+// newBusyboxBin returns a [pkg.Artifact] containing a busybox installation from
+// the https://busybox.net/downloads/binaries/ binary release.
+func newBusyboxBin() pkg.Artifact {
+	const (
+		version  = "1.35.0"
+		checksum = "L7OBIsPu9enNHn7FqpBT1kOg_mCLNmetSeNMA3i4Y60Z5jTgnlX3qX3zcQtLx5AB"
+	)
+	return pkg.NewExec(
+		"busybox-bin-"+version, nil, pkg.ExecTimeoutMax, false,
+		fhs.AbsRoot, []string{
+			"PATH=/system/bin",
+		},
+		AbsSystem.Append("bin", "busybox"),
+		[]string{"hush", "-c", "" +
+			"busybox mkdir -p /work/system/bin/ && " +
+			"busybox cp /system/bin/busybox /work/system/bin/ && " +
+			"busybox --install -s /work/system/bin/"},
+		pkg.Path(AbsSystem, true, busyboxBin{pkg.NewHTTPGet(
+			&http.Client{Transport: &http.Transport{
+				// busybox website is really slow to respond
+				TLSHandshakeTimeout: 2 * time.Minute,
+			}},
+			"https://busybox.net/downloads/binaries/"+
+				version+"-"+linuxArch()+"-linux-musl/busybox",
+			mustDecode(checksum),
+		)}),
+	)
+}
+
+func (t Toolchain) newBusybox() pkg.Artifact {
+	const (
+		version  = "1.37.0"
+		checksum = "Ial94Tnt7esJ_YEeb0AxunVL6MGYFyOw7Rtu2o87CXCi1TLrc6rlznVsN1rZk7it"
+	)
+
+	var env []string
+	if t == toolchainStage3 {
+		env = append(env, "EXTRA_LDFLAGS=-static")
+	}
+
+	return t.New("busybox-"+version, false, stage3Concat(t, []pkg.Artifact{},
+		t.Load(Make),
+		t.Load(KernelHeaders),
+	), nil, slices.Concat([]string{
+		"ROSA_BUSYBOX_ENABLE=" + strings.Join([]string{
+			"STATIC",
+			"PIE",
+		}, " "),
+		"ROSA_BUSYBOX_DISABLE=" + strings.Join([]string{
+			"FEATURE_IPV6",
+			"FEATURE_PREFER_IPV4_ADDRESS",
+			"FEATURE_HWIB",
+			"ARP",
+			"ARPING",
+			"BRCTL",
+			"FEATURE_BRCTL_FANCY",
+			"FEATURE_BRCTL_SHOW",
+			"DNSD",
+			"ETHER_WAKE",
+			"FTPD",
+			"FEATURE_FTPD_WRITE",
+			"FEATURE_FTPD_ACCEPT_BROKEN_LIST",
+			"FEATURE_FTPD_AUTHENTICATION",
+			"FTPGET",
+			"FTPPUT",
+			"FEATURE_FTPGETPUT_LONG_OPTIONS",
+			"HOSTNAME",
+			"DNSDOMAINNAME",
+			"HTTPD",
+			"FEATURE_HTTPD_PORT_DEFAULT",
+			"FEATURE_HTTPD_RANGES",
+			"FEATURE_HTTPD_SETUID",
+			"FEATURE_HTTPD_BASIC_AUTH",
+			"FEATURE_HTTPD_AUTH_MD5",
+			"FEATURE_HTTPD_CGI",
+			"FEATURE_HTTPD_CONFIG_WITH_SCRIPT_INTERPR",
+			"FEATURE_HTTPD_SET_REMOTE_PORT_TO_ENV",
+			"FEATURE_HTTPD_ENCODE_URL_STR",
+			"FEATURE_HTTPD_ERROR_PAGES",
+			"FEATURE_HTTPD_PROXY",
+			"FEATURE_HTTPD_GZIP",
+			"FEATURE_HTTPD_ETAG",
+			"FEATURE_HTTPD_LAST_MODIFIED",
+			"FEATURE_HTTPD_DATE",
+			"FEATURE_HTTPD_ACL_IP",
+			"IFCONFIG",
+			"FEATURE_IFCONFIG_STATUS",
+			"FEATURE_IFCONFIG_SLIP",
+			"FEATURE_IFCONFIG_MEMSTART_IOADDR_IRQ",
+			"FEATURE_IFCONFIG_HW",
+			"FEATURE_IFCONFIG_BROADCAST_PLUS",
+			"IFENSLAVE",
+			"IFPLUGD",
+			"IFUP",
+			"IFDOWN",
+			"IFUPDOWN_IFSTATE_PATH",
+			"FEATURE_IFUPDOWN_IP",
+			"FEATURE_IFUPDOWN_IPV4",
+			"FEATURE_IFUPDOWN_IPV6",
+			"FEATURE_IFUPDOWN_MAPPING",
+			"INETD",
+			"FEATURE_INETD_SUPPORT_BUILTIN_ECHO",
+			"FEATURE_INETD_SUPPORT_BUILTIN_DISCARD",
+			"FEATURE_INETD_SUPPORT_BUILTIN_TIME",
+			"FEATURE_INETD_SUPPORT_BUILTIN_DAYTIME",
+			"FEATURE_INETD_SUPPORT_BUILTIN_CHARGEN",
+			"IP",
+			"IPADDR",
+			"IPLINK",
+			"IPROUTE",
+			"IPTUNNEL",
+			"IPRULE",
+			"IPNEIGH",
+			"FEATURE_IP_ADDRESS",
+			"FEATURE_IP_LINK",
+			"FEATURE_IP_LINK_CAN",
+			"FEATURE_IP_ROUTE",
+			"FEATURE_IP_ROUTE_DIR",
+			"FEATURE_IP_TUNNEL",
+			"FEATURE_IP_RULE",
+			"FEATURE_IP_NEIGH",
+			"IPCALC",
+			"FEATURE_IPCALC_LONG_OPTIONS",
+			"FEATURE_IPCALC_FANCY",
+			"FAKEIDENTD",
+			"NAMEIF",
+			"FEATURE_NAMEIF_EXTENDED",
+			"NBDCLIENT",
+			"NC",
+			"NC_SERVER",
+			"NC_EXTRA",
+			"NC_110_COMPAT",
+			"NETSTAT",
+			"FEATURE_NETSTAT_WIDE",
+			"FEATURE_NETSTAT_PRG",
+			"NSLOOKUP",
+			"FEATURE_NSLOOKUP_BIG",
+			"FEATURE_NSLOOKUP_LONG_OPTIONS",
+			"NTPD",
+			"FEATURE_NTPD_SERVER",
+			"FEATURE_NTPD_CONF",
+			"FEATURE_NTP_AUTH",
+			"PING",
+			"PING6",
+			"FEATURE_FANCY_PING",
+			"PSCAN",
+			"ROUTE",
+			"SLATTACH",
+			"SSL_CLIENT",
+			"TC",
+			"FEATURE_TC_INGRESS",
+			"TCPSVD",
+			"UDPSVD",
+			"TELNET",
+			"FEATURE_TELNET_TTYPE",
+			"FEATURE_TELNET_AUTOLOGIN",
+			"FEATURE_TELNET_WIDTH",
+			"TELNETD",
+			"FEATURE_TELNETD_STANDALONE",
+			"FEATURE_TELNETD_PORT_DEFAULT",
+			"FEATURE_TELNETD_INETD_WAIT",
+			"TFTP",
+			"FEATURE_TFTP_PROGRESS_BAR",
+			"FEATURE_TFTP_HPA_COMPAT",
+			"TFTPD",
+			"FEATURE_TFTP_GET",
+			"FEATURE_TFTP_PUT",
+			"FEATURE_TFTP_BLOCKSIZE",
+			"TLS",
+			"TRACEROUTE",
+			"TRACEROUTE6",
+			"FEATURE_TRACEROUTE_VERBOSE",
+			"FEATURE_TRACEROUTE_USE_ICMP",
+			"TUNCTL",
+			"FEATURE_TUNCTL_UG",
+			"VCONFIG",
+			"WGET",
+			"FEATURE_WGET_LONG_OPTIONS",
+			"FEATURE_WGET_STATUSBAR",
+			"FEATURE_WGET_FTP",
+			"FEATURE_WGET_AUTHENTICATION",
+			"FEATURE_WGET_TIMEOUT",
+			"FEATURE_WGET_HTTPS",
+			"FEATURE_WGET_OPENSSL",
+			"WHOIS",
+			"ZCIP",
+			"UDHCPD",
+			"FEATURE_UDHCPD_BOOTP",
+			"FEATURE_UDHCPD_WRITE_LEASES_EARLY",
+			"DHCPD_LEASES_FILE",
+			"DUMPLEASES",
+			"DHCPRELAY",
+			"UDHCPC",
+			"FEATURE_UDHCPC_ARPING",
+			"FEATURE_UDHCPC_SANITIZEOPT",
+			"UDHCPC_DEFAULT_SCRIPT",
+			"UDHCPC6_DEFAULT_SCRIPT",
+			"UDHCPC6",
+			"FEATURE_UDHCPC6_RFC3646",
+			"FEATURE_UDHCPC6_RFC4704",
+			"FEATURE_UDHCPC6_RFC4833",
+			"FEATURE_UDHCPC6_RFC5970",
+		}, " "),
+	}, env), `
+config_enable() {
+	for ent in "$@"; do
+		sed "s/^# CONFIG_${ent}.*/CONFIG_${ent}=y/" -i .config
+		shift
+	done
+}
+
+config_disable() {
+	for ent in "$@"; do
+		sed "s/^CONFIG_${ent}=y/# CONFIG_${ent} is not set/" -i .config
+		shift
+	done
+}
+
+cat > /bin/gcc << EOF
+exec clang \
+	-Wno-ignored-optimization-argument \
+	${LDFLAGS} \
+	\$@
+EOF
+chmod +x /bin/gcc
+
+cd /usr/src/busybox
+chmod +w editors editors/awk.c
+patch -p 1 < /usr/src/patches/awk-fix-literal-backslash.patch
+
+cd "$(mktemp -d)"
+make \
+	KBUILD_SRC=/usr/src/busybox \
+	-f /usr/src/busybox/Makefile \
+	defconfig
+
+config_enable $ROSA_BUSYBOX_ENABLE
+config_disable $ROSA_BUSYBOX_DISABLE
+ln -s ../system/bin/pwd /bin/pwd || true
+make CFLAGS_busybox="${LDFLAGS} ${EXTRA_LDFLAGS}" "-j$(nproc)"
+
+mkdir -p /system/bin/ /work/bin/
+cp busybox /system/bin/
+
+mkdir -pv /work/system/bin/
+busybox --install -s /work/system/bin/
+cp -v busybox /work/system/bin/
+ln -vs ../system/bin/hush /work/bin/sh
+mkdir -vp /work/usr/bin/
+ln -vs ../../system/bin/busybox /work/usr/bin/env
+`, pkg.Path(AbsUsrSrc.Append("busybox"), true, pkg.NewHTTPGetTar(
+		&http.Client{Transport: &http.Transport{
+			// busybox website is really slow to respond
+			TLSHandshakeTimeout: 2 * time.Minute,
+		}},
+		"https://busybox.net/downloads/busybox-"+version+".tar.bz2",
+		mustDecode(checksum),
+		pkg.TarBzip2,
+	)), pkg.Path(
+		AbsUsrSrc.Append("patches", "awk-fix-literal-backslash.patch"), false,
+		pkg.NewFile("awk-fix-literal-backslash.patch", []byte(`diff --git a/editors/awk.c b/editors/awk.c
+index 64e752f4b..40f5ba7f7 100644
+--- a/editors/awk.c
++++ b/editors/awk.c
+@@ -2636,8 +2636,13 @@ static int awk_sub(node *rn, const char *repl, int nm, var *src, var *dest /*,in
+ 					resbuf = qrealloc(resbuf, residx + replen + n, &resbufsize);
+ 					memcpy(resbuf + residx, sp + pmatch[j].rm_so - start_ofs, n);
+ 					residx += n;
+-				} else
++				} else {
++/* '\\' and '&' following a backslash keep its original meaning, any other
++ * occurrence of a '\\' should be treated as literal */
++					if (bslash && c != '\\' && c != '&')
++						resbuf[residx++] = '\\';
+ 					resbuf[residx++] = c;
++				}
+ 				bslash = 0;
+ 			}
+ 		}`)),
+	))
+}
+func init() { artifactsF[Busybox] = Toolchain.newBusybox }

internal/rosa/cmake.go

@@ -0,0 +1,125 @@
+package rosa
+
+import (
+	"slices"
+	"strings"
+
+	"hakurei.app/container/check"
+	"hakurei.app/internal/pkg"
+)
+
+func (t Toolchain) newCMake() pkg.Artifact {
+	const (
+		version  = "4.2.1"
+		checksum = "Y3OdbMsob6Xk2y1DCME6z4Fryb5_TkFD7knRT8dTNIRtSqbiCJyyDN9AxggN_I75"
+	)
+	return t.New("cmake-"+version, false, []pkg.Artifact{
+		t.Load(Make),
+		t.Load(KernelHeaders),
+	}, nil, nil, `
+# expected to be writable in the copy made during bootstrap
+chmod -R +w /usr/src/cmake/Tests
+
+cd "$(mktemp -d)"
+/usr/src/cmake/bootstrap \
+	--prefix=/system \
+	--parallel="$(nproc)" \
+	-- \
+	-DCMAKE_USE_OPENSSL=OFF
+make "-j$(nproc)"
+make DESTDIR=/work install
+`, pkg.Path(AbsUsrSrc.Append("cmake"), true,
+		pkg.NewHTTPGetTar(
+			nil, "https://github.com/Kitware/CMake/releases/download/"+
+				"v"+version+"/cmake-"+version+".tar.gz",
+			mustDecode(checksum),
+			pkg.TarGzip,
+		)))
+}
+func init() { artifactsF[CMake] = Toolchain.newCMake }
+
+// CMakeAttr holds the project-specific attributes that will be applied to a new
+// [pkg.Artifact] compiled via CMake.
+type CMakeAttr struct {
+	// Path elements joined with source.
+	Append []string
+	// Use source tree as scratch space.
+	Writable bool
+	// Dependencies concatenated with the build system itself.
+	Extra []pkg.Artifact
+
+	// CMake CACHE entries.
+	Cache [][2]string
+	// Additional environment variables.
+	Env []string
+	// Runs before cmake.
+	ScriptEarly string
+	// Runs after cmake.
+	Script string
+
+	// Override the default installation prefix [AbsSystem].
+	Prefix *check.Absolute
+
+	// Return an exclusive artifact.
+	Exclusive bool
+}
+
+// NewViaCMake returns a [pkg.Artifact] for compiling and installing via CMake.
+func (t Toolchain) NewViaCMake(
+	name, version, variant string,
+	source pkg.Artifact,
+	attr *CMakeAttr,
+) pkg.Artifact {
+	if name == "" || version == "" || variant == "" {
+		panic("names must be non-empty")
+	}
+	if attr == nil {
+		attr = &CMakeAttr{
+			Cache: [][2]string{
+				{"CMAKE_BUILD_TYPE", "Release"},
+			},
+		}
+	}
+	if len(attr.Cache) == 0 {
+		panic("CACHE must be non-empty")
+	}
+
+	scriptEarly := attr.ScriptEarly
+	if attr.Writable {
+		scriptEarly = `
+chmod -R +w "${ROSA_SOURCE}"
+` + scriptEarly
+	}
+
+	prefix := attr.Prefix
+	if prefix == nil {
+		prefix = AbsSystem
+	}
+
+	sourcePath := AbsUsrSrc.Append(name)
+	return t.New(name+"-"+variant+"-"+version, attr.Exclusive, stage3Concat(t, attr.Extra,
+		t.Load(CMake),
+		t.Load(Ninja),
+	), nil, slices.Concat([]string{
+		"ROSA_SOURCE=" + sourcePath.String(),
+		"ROSA_CMAKE_SOURCE=" + sourcePath.Append(attr.Append...).String(),
+		"ROSA_INSTALL_PREFIX=/work" + prefix.String(),
+	}, attr.Env), scriptEarly+`
+mkdir /cure && cd /cure
+cmake -G Ninja \
+	-DCMAKE_C_COMPILER_TARGET="${ROSA_TRIPLE}" \
+	-DCMAKE_CXX_COMPILER_TARGET="${ROSA_TRIPLE}" \
+	-DCMAKE_ASM_COMPILER_TARGET="${ROSA_TRIPLE}" \
+	`+strings.Join(slices.Collect(func(yield func(string) bool) {
+		for _, v := range attr.Cache {
+			if !yield("-D" + v[0] + "=" + v[1]) {
+				return
+			}
+		}
+	}), " \\\n\t")+` \
+	-DCMAKE_INSTALL_PREFIX="${ROSA_INSTALL_PREFIX}" \
+	"${ROSA_CMAKE_SOURCE}"
+cmake --build .
+cmake --install .
+`+attr.Script, pkg.Path(sourcePath, attr.Writable, source))
+}