cmd/irdump: formatted disassembly

.gitea/workflows/test.yml

@@ -89,6 +89,23 @@ jobs:
           path: result/*
           retention-days: 1
 
+  hpkg:
+    name: Hpkg
+    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.hpkg
+
+      - name: Upload test output
+        uses: actions/upload-artifact@v3
+        with:
+          name: "hpkg-vm-output"
+          path: result/*
+          retention-days: 1
+
   check:
     name: Flake checks
     needs:
@@ -97,6 +114,7 @@ jobs:
       - sandbox
       - sandbox-race
       - sharefs
+      - hpkg
     runs-on: nix
     steps:
       - name: Checkout

.github/workflows/README

@@ -0,0 +1,5 @@
+DO NOT ADD NEW ACTIONS HERE
+
+This port is solely for releasing to the github mirror and serves no purpose during development.
+All development happens at https://git.gensokyo.uk/security/hakurei. If you wish to contribute,
+request for an account on git.gensokyo.uk.

.github/workflows/release.yml

@@ -0,0 +1,46 @@
+name: Release
+
+on:
+  push:
+    tags:
+      - 'v*'
+
+jobs:
+  release:
+    name: Create release
+    runs-on: ubuntu-latest
+
+    permissions:
+      packages: write
+      contents: write
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Install Nix
+        uses: nixbuild/nix-quick-install-action@v32
+        with:
+          nix_conf: |
+            keep-env-derivations = true
+            keep-outputs = true
+
+      - name: Restore and cache Nix store
+        uses: nix-community/cache-nix-action@v6
+        with:
+          primary-key: build-${{ runner.os }}-${{ hashFiles('**/*.nix') }}
+          restore-prefixes-first-match: build-${{ runner.os }}-
+          gc-max-store-size-linux: 1G
+          purge: true
+          purge-prefixes: build-${{ runner.os }}-
+          purge-created: 60
+          purge-primary-key: never
+
+      - name: Build for release
+        run: nix build --print-out-paths --print-build-logs .#dist
+
+      - name: Release
+        uses: softprops/action-gh-release@v2
+        with:
+          files: |-
+            result/hakurei-**

.github/workflows/test.yml

@@ -0,0 +1,48 @@
+name: Test
+
+on:
+  - push
+
+jobs:
+  dist:
+    name: Create distribution
+    runs-on: ubuntu-latest
+    permissions:
+      actions: write
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Install Nix
+        uses: nixbuild/nix-quick-install-action@v32
+        with:
+          nix_conf: |
+            keep-env-derivations = true
+            keep-outputs = true
+
+      - name: Restore and cache Nix store
+        uses: nix-community/cache-nix-action@v6
+        with:
+          primary-key: build-${{ runner.os }}-${{ hashFiles('**/*.nix') }}
+          restore-prefixes-first-match: build-${{ runner.os }}-
+          gc-max-store-size-linux: 1G
+          purge: true
+          purge-prefixes: build-${{ runner.os }}-
+          purge-created: 60
+          purge-primary-key: never
+
+      - name: Build for test
+        id: build-test
+        run: >-
+          export HAKUREI_REV="$(git rev-parse --short HEAD)" &&
+          sed -i.old 's/version = /version = "0.0.0-'$HAKUREI_REV'"; # version = /' package.nix &&
+          nix build --print-out-paths --print-build-logs .#dist &&
+          mv package.nix.old package.nix &&
+          echo "rev=$HAKUREI_REV" >> $GITHUB_OUTPUT
+
+      - name: Upload test build
+        uses: actions/upload-artifact@v4
+        with:
+          name: "hakurei-${{ steps.build-test.outputs.rev }}"
+          path: result/*
+          retention-days: 1

.gitignore

@@ -1,14 +1,36 @@
-# produced by tools and text editors
+# Binaries for programs and plugins
-*.qcow2
+*.exe
+*.exe~
+*.dll
+*.so
+*.dylib
+*.pkg
+/hakurei
+
+# Test binary, built with `go test -c`
 *.test
+
+# Output of the go coverage tool, specifically when used with LiteIDE
 *.out
+
+# Dependency directories (remove the comment below to include it)
+# vendor/
+
+# Go workspace file
+go.work
+go.work.sum
+
+# env file
+.env
 .idea
 .vscode
 
 # go generate
 /cmd/hakurei/LICENSE
 /internal/pkg/testdata/testtool
-/internal/rosa/hakurei_current.tar.gz
 
-# cmd/dist default destination
+# release
-/dist
+/dist/hakurei-*
+
+# interactive nixos vm
+nixos.qcow2

README.md

@@ -1,5 +1,5 @@
 <p align="center">
-  <a href="https://git.gensokyo.uk/rosa/hakurei">
+  <a href="https://git.gensokyo.uk/security/hakurei">
     <picture>
       <img src="https://basement.gensokyo.uk/images/yukari1.png" width="200px" alt="Yukari">
     </picture>
@@ -8,58 +8,171 @@
 
 <p align="center">
   <a href="https://pkg.go.dev/hakurei.app"><img src="https://pkg.go.dev/badge/hakurei.app.svg" alt="Go Reference" /></a>
-  <a href="https://git.gensokyo.uk/rosa/hakurei/actions"><img src="https://git.gensokyo.uk/rosa/hakurei/actions/workflows/test.yml/badge.svg?branch=staging&style=flat-square" alt="Gitea Workflow Status" /></a>
+  <a href="https://git.gensokyo.uk/security/hakurei/actions"><img src="https://git.gensokyo.uk/security/hakurei/actions/workflows/test.yml/badge.svg?branch=staging&style=flat-square" alt="Gitea Workflow Status" /></a>
   <br/>
-  <a href="https://git.gensokyo.uk/rosa/hakurei/releases"><img src="https://img.shields.io/gitea/v/release/rosa/hakurei?gitea_url=https%3A%2F%2Fgit.gensokyo.uk&color=purple" alt="Release" /></a>
+  <a href="https://git.gensokyo.uk/security/hakurei/releases"><img src="https://img.shields.io/gitea/v/release/security/hakurei?gitea_url=https%3A%2F%2Fgit.gensokyo.uk&color=purple" alt="Release" /></a>
   <a href="https://goreportcard.com/report/hakurei.app"><img src="https://goreportcard.com/badge/hakurei.app" alt="Go Report Card" /></a>
   <a href="https://hakurei.app"><img src="https://img.shields.io/website?url=https%3A%2F%2Fhakurei.app" alt="Website" /></a>
 </p>
 
-Hakurei is a tool for running sandboxed desktop applications as dedicated
+Hakurei is a tool for running sandboxed graphical applications as dedicated subordinate users on the Linux kernel.
-subordinate users on the Linux kernel. It implements the application container
+It implements the application container of [planterette (WIP)](https://git.gensokyo.uk/security/planterette),
-of [planterette (WIP)](https://git.gensokyo.uk/rosa/planterette), a
+a self-contained Android-like package manager with modern security features.
-self-contained Android-like package manager with modern security features.
 
-Interaction with hakurei happens entirely through structures described by
+## NixOS Module usage
-package [hst](https://pkg.go.dev/hakurei.app/hst). No native API is available
-due to internal details of uid isolation.
 
-## Notable Packages
+The NixOS module currently requires home-manager to configure subordinate users. Full module documentation can be found [here](options.md).
 
-Package [container](https://pkg.go.dev/hakurei.app/container) is general purpose
+To use the module, import it into your configuration with
-container tooling. It is used by the hakurei shim process running as the target
-subordinate user to set up the application container. It has a single dependency,
-[libseccomp](https://github.com/seccomp/libseccomp), to create BPF programs
-for the [system call filter](https://www.kernel.org/doc/html/latest/userspace-api/seccomp_filter.html).
 
-Package [internal/pkg](https://pkg.go.dev/hakurei.app/internal/pkg) provides
+```nix
-infrastructure for hermetic builds. This replaces the legacy nix-based testing
+{
-framework and serves as the build system of Rosa OS, currently developed under
+  inputs = {
-package [internal/rosa](https://pkg.go.dev/hakurei.app/internal/rosa).
+    nixpkgs.url = "github:NixOS/nixpkgs/nixos-24.11";
 
-## Dependencies
+    hakurei = {
+      url = "git+https://git.gensokyo.uk/security/hakurei";
 
-`container` depends on:
+      # Optional but recommended to limit the size of your system closure.
+      inputs.nixpkgs.follows = "nixpkgs";
+    };
+  };
 
-- [libseccomp](https://github.com/seccomp/libseccomp) to generate BPF programs.
+  outputs = { self, nixpkgs, hakurei, ... }:
+  {
+    nixosConfigurations.hakurei = nixpkgs.lib.nixosSystem {
+      system = "x86_64-linux";
+      modules = [
+        hakurei.nixosModules.hakurei
+      ];
+    };
+  };
+}
+```
 
-`cmd/hakurei` depends on:
+This adds the `environment.hakurei` option:
 
-- [acl](https://savannah.nongnu.org/projects/acl/) to export sockets to
+```nix
-  subordinate users.
+{ pkgs, ... }:
-- [wayland](https://gitlab.freedesktop.org/wayland/wayland) to set up
-  [security-context-v1](https://wayland.app/protocols/security-context-v1).
-- [xcb](https://xcb.freedesktop.org/) to grant and revoke subordinate users
-  access to the X server.
 
-`cmd/sharefs` depends on:
+{
+  environment.hakurei = {
+    enable = true;
+    stateDir = "/var/lib/hakurei";
+    users = {
+      alice = 0;
+      nixos = 10;
+    };
 
-- [fuse](https://github.com/libfuse/libfuse) to implement the filesystem.
+    commonPaths = [
+      {
+        src = "/sdcard";
+        write = true;
+      }
+    ];
 
-New dependencies will generally not be added. Patches adding new dependencies
+    extraHomeConfig = {
-are very likely to be rejected.
+      home.stateVersion = "23.05";
+    };
 
-## NixOS Module (deprecated)
+    apps = {
+      "org.chromium.Chromium" = {
+        name = "chromium";
+        identity = 1;
+        packages = [ pkgs.chromium ];
+        userns = true;
+        mapRealUid = true;
+        dbus = {
+          system = {
+            filter = true;
+            talk = [
+              "org.bluez"
+              "org.freedesktop.Avahi"
+              "org.freedesktop.UPower"
+            ];
+          };
+          session =
+            f:
+            f {
+              talk = [
+                "org.freedesktop.FileManager1"
+                "org.freedesktop.Notifications"
+                "org.freedesktop.ScreenSaver"
+                "org.freedesktop.secrets"
+                "org.kde.kwalletd5"
+                "org.kde.kwalletd6"
+              ];
+              own = [
+                "org.chromium.Chromium.*"
+                "org.mpris.MediaPlayer2.org.chromium.Chromium.*"
+                "org.mpris.MediaPlayer2.chromium.*"
+              ];
+              call = { };
+              broadcast = { };
+            };
+        };
+      };
 
-The NixOS module is in maintenance mode and will be removed once planterette is
+      "org.claws_mail.Claws-Mail" = {
-feature-complete. Full module documentation can be found [here](options.md).
+        name = "claws-mail";
+        identity = 2;
+        packages = [ pkgs.claws-mail ];
+        gpu = false;
+        capability.pulse = false;
+      };
+
+      "org.weechat" = {
+        name = "weechat";
+        identity = 3;
+        shareUid = true;
+        packages = [ pkgs.weechat ];
+        capability = {
+          wayland = false;
+          x11 = false;
+          dbus = true;
+          pulse = false;
+        };
+      };
+
+      "dev.vencord.Vesktop" = {
+        name = "discord";
+        identity = 3;
+        shareUid = true;
+        packages = [ pkgs.vesktop ];
+        share = pkgs.vesktop;
+        command = "vesktop --ozone-platform-hint=wayland";
+        userns = true;
+        mapRealUid = true;
+        capability.x11 = true;
+        dbus = {
+          session =
+            f:
+            f {
+              talk = [ "org.kde.StatusNotifierWatcher" ];
+              own = [ ];
+              call = { };
+              broadcast = { };
+            };
+          system.filter = true;
+        };
+      };
+
+      "io.looking-glass" = {
+        name = "looking-glass-client";
+        identity = 4;
+        useCommonPaths = false;
+        groups = [ "plugdev" ];
+        extraPaths = [
+          {
+            src = "/dev/shm/looking-glass";
+            write = true;
+          }
+        ];
+        extraConfig = {
+          programs.looking-glass-client.enable = true;
+        };
+      };
+    };
+  };
+}
+```

all.sh

@@ -1,6 +0,0 @@
-#!/bin/sh -e
-
-TOOLCHAIN_VERSION="$(go version)"
-cd "$(dirname -- "$0")/"
-echo "# Building cmd/dist using ${TOOLCHAIN_VERSION}."
-go run -v --tags=dist ./cmd/dist

cmd/dist/main.go

@@ -1,237 +0,0 @@
-//go:build dist
-
-package main
-
-import (
-	"archive/tar"
-	"compress/gzip"
-	"context"
-	"crypto/sha512"
-	_ "embed"
-	"encoding/hex"
-	"fmt"
-	"io"
-	"io/fs"
-	"log"
-	"os"
-	"os/exec"
-	"os/signal"
-	"path/filepath"
-	"runtime"
-)
-
-// getenv looks up an environment variable, and returns fallback if it is unset.
-func getenv(key, fallback string) string {
-	if v, ok := os.LookupEnv(key); ok {
-		return v
-	}
-	return fallback
-}
-
-// mustRun runs a command with the current process's environment and panics
-// on error or non-zero exit code.
-func mustRun(ctx context.Context, name string, arg ...string) {
-	cmd := exec.CommandContext(ctx, name, arg...)
-	cmd.Stdin, cmd.Stdout, cmd.Stderr = os.Stdin, os.Stdout, os.Stderr
-	if err := cmd.Run(); err != nil {
-		panic(err)
-	}
-}
-
-//go:embed comp/_hakurei
-var comp []byte
-
-func main() {
-	fmt.Println()
-	log.SetFlags(0)
-	log.SetPrefix("# ")
-
-	version := getenv("HAKUREI_VERSION", "untagged")
-	prefix := getenv("PREFIX", "/usr")
-	destdir := getenv("DESTDIR", "dist")
-
-	if err := os.MkdirAll(destdir, 0755); err != nil {
-		log.Fatal(err)
-	}
-	s, err := os.MkdirTemp(destdir, ".dist.*")
-	if err != nil {
-		log.Fatal(err)
-	}
-	defer func() {
-		var code int
-
-		if err = os.RemoveAll(s); err != nil {
-			code = 1
-			log.Println(err)
-		}
-
-		if r := recover(); r != nil {
-			code = 1
-			log.Println(r)
-		}
-
-		os.Exit(code)
-	}()
-
-	ctx, cancel := signal.NotifyContext(context.Background(), os.Interrupt)
-	defer cancel()
-
-	log.Println("Building hakurei.")
-	mustRun(ctx, "go", "generate", "./...")
-	mustRun(
-		ctx, "go", "build",
-		"-trimpath",
-		"-v", "-o", s,
-		"-ldflags=-s -w "+
-			"-buildid= -linkmode external -extldflags=-static "+
-			"-X hakurei.app/internal/info.buildVersion="+version+" "+
-			"-X hakurei.app/internal/info.hakureiPath="+prefix+"/bin/hakurei "+
-			"-X hakurei.app/internal/info.hsuPath="+prefix+"/bin/hsu "+
-			"-X main.hakureiPath="+prefix+"/bin/hakurei",
-		"./...",
-	)
-	fmt.Println()
-
-	log.Println("Testing Hakurei.")
-	mustRun(
-		ctx, "go", "test",
-		"-ldflags=-buildid= -linkmode external -extldflags=-static",
-		"./...",
-	)
-	fmt.Println()
-
-	log.Println("Creating distribution.")
-	const suffix = ".tar.gz"
-	distName := "hakurei-" + version + "-" + runtime.GOARCH
-	var f *os.File
-	if f, err = os.OpenFile(
-		filepath.Join(s, distName+suffix),
-		os.O_CREATE|os.O_EXCL|os.O_WRONLY,
-		0644,
-	); err != nil {
-		panic(err)
-	}
-	defer func() {
-		if f == nil {
-			return
-		}
-		if err = f.Close(); err != nil {
-			log.Println(err)
-		}
-	}()
-
-	h := sha512.New()
-	gw := gzip.NewWriter(io.MultiWriter(f, h))
-	tw := tar.NewWriter(gw)
-
-	mustWriteHeader := func(name string, size int64, mode os.FileMode) {
-		header := tar.Header{
-			Name:  filepath.Join(distName, name),
-			Size:  size,
-			Mode:  int64(mode),
-			Uname: "root",
-			Gname: "root",
-		}
-
-		if mode&os.ModeDir != 0 {
-			header.Typeflag = tar.TypeDir
-			fmt.Printf("%s %s\n", mode, name)
-		} else {
-			header.Typeflag = tar.TypeReg
-			fmt.Printf("%s %s (%d bytes)\n", mode, name, size)
-		}
-
-		if err = tw.WriteHeader(&header); err != nil {
-			panic(err)
-		}
-	}
-
-	mustWriteFile := func(name string, data []byte, mode os.FileMode) {
-		mustWriteHeader(name, int64(len(data)), mode)
-		if mode&os.ModeDir != 0 {
-			return
-		}
-		if _, err = tw.Write(data); err != nil {
-			panic(err)
-		}
-	}
-
-	mustWriteFromPath := func(dst, src string, mode os.FileMode) {
-		var r *os.File
-		if r, err = os.Open(src); err != nil {
-			panic(err)
-		}
-
-		var fi os.FileInfo
-		if fi, err = r.Stat(); err != nil {
-			_ = r.Close()
-			panic(err)
-		}
-
-		if mode == 0 {
-			mode = fi.Mode()
-		}
-
-		mustWriteHeader(dst, fi.Size(), mode)
-		if _, err = io.Copy(tw, r); err != nil {
-			_ = r.Close()
-			panic(err)
-		} else if err = r.Close(); err != nil {
-			panic(err)
-		}
-	}
-
-	mustWriteFile(".", nil, fs.ModeDir|0755)
-	mustWriteFile("comp/", nil, os.ModeDir|0755)
-	mustWriteFile("comp/_hakurei", comp, 0644)
-	mustWriteFile("install.sh", []byte(`#!/bin/sh -e
-cd "$(dirname -- "$0")" || exit 1
-
-install -vDm0755 "bin/hakurei" "${DESTDIR}`+prefix+`/bin/hakurei"
-install -vDm0755 "bin/sharefs" "${DESTDIR}`+prefix+`/bin/sharefs"
-
-install -vDm4511 "bin/hsu" "${DESTDIR}`+prefix+`/bin/hsu"
-if [ ! -f "${DESTDIR}/etc/hsurc" ]; then
-    install -vDm0400 "hsurc.default" "${DESTDIR}/etc/hsurc"
-fi
-
-install -vDm0644 "comp/_hakurei" "${DESTDIR}`+prefix+`/share/zsh/site-functions/_hakurei"
-`), 0755)
-
-	mustWriteFromPath("README.md", "README.md", 0)
-	mustWriteFile("hsurc.default", []byte("1000 0"), 0400)
-	mustWriteFromPath("bin/hsu", filepath.Join(s, "hsu"), 04511)
-	for _, name := range []string{
-		"hakurei",
-		"sharefs",
-	} {
-		mustWriteFromPath(
-			filepath.Join("bin", name),
-			filepath.Join(s, name),
-			0,
-		)
-	}
-
-	if err = tw.Close(); err != nil {
-		panic(err)
-	} else if err = gw.Close(); err != nil {
-		panic(err)
-	} else if err = f.Close(); err != nil {
-		panic(err)
-	}
-	f = nil
-
-	if err = os.WriteFile(
-		filepath.Join(destdir, distName+suffix+".sha512"),
-		append(hex.AppendEncode(nil, h.Sum(nil)), "  "+distName+suffix+"\n"...),
-		0644,
-	); err != nil {
-		panic(err)
-	}
-	if err = os.Rename(
-		filepath.Join(s, distName+suffix),
-		filepath.Join(destdir, distName+suffix),
-	); err != nil {
-		panic(err)
-	}
-}

cmd/earlyinit/main.go

@@ -1,131 +0,0 @@
-// The earlyinit is part of the Rosa OS initramfs and serves as the system init.
-//
-// This program is an internal detail of Rosa OS and is not usable on its own.
-// It is not covered by the compatibility promise.
-package main
-
-import (
-	"log"
-	"os"
-	"runtime"
-	"strings"
-	. "syscall"
-)
-
-func main() {
-	runtime.LockOSThread()
-	log.SetFlags(0)
-	log.SetPrefix("earlyinit: ")
-
-	var (
-		option map[string]string
-		flags  []string
-	)
-	if len(os.Args) > 1 {
-		option = make(map[string]string)
-		for _, s := range os.Args[1:] {
-			key, value, ok := strings.Cut(s, "=")
-			if !ok {
-				flags = append(flags, s)
-				continue
-			}
-			option[key] = value
-		}
-	}
-
-	if err := Mount(
-		"devtmpfs",
-		"/dev/",
-		"devtmpfs",
-		MS_NOSUID|MS_NOEXEC,
-		"",
-	); err != nil {
-		log.Fatalf("cannot mount devtmpfs: %v", err)
-	}
-
-	// The kernel might be unable to set up the console. When that happens,
-	// printk is called with "Warning: unable to open an initial console."
-	// and the init runs with no files. The checkfds runtime function
-	// populates 0-2 by opening /dev/null for them.
-	//
-	// This check replaces 1 and 2 with /dev/kmsg to improve the chance
-	// of output being visible to the user.
-	if fi, err := os.Stdout.Stat(); err == nil {
-		if stat, ok := fi.Sys().(*Stat_t); ok {
-			if stat.Rdev == 0x103 {
-				var fd int
-				if fd, err = Open(
-					"/dev/kmsg",
-					O_WRONLY|O_CLOEXEC,
-					0,
-				); err != nil {
-					log.Fatalf("cannot open kmsg: %v", err)
-				}
-
-				if err = Dup3(fd, Stdout, 0); err != nil {
-					log.Fatalf("cannot open stdout: %v", err)
-				}
-				if err = Dup3(fd, Stderr, 0); err != nil {
-					log.Fatalf("cannot open stderr: %v", err)
-				}
-
-				if err = Close(fd); err != nil {
-					log.Printf("cannot close kmsg: %v", err)
-				}
-			}
-		}
-	}
-
-	// staying in rootfs, these are no longer used
-	must(os.Remove("/root"))
-	must(os.Remove("/init"))
-
-	must(os.Mkdir("/proc", 0))
-	mustSyscall("mount proc", Mount(
-		"proc",
-		"/proc",
-		"proc",
-		MS_NOSUID|MS_NOEXEC|MS_NODEV,
-		"hidepid=1",
-	))
-
-	must(os.Mkdir("/sys", 0))
-	mustSyscall("mount sysfs", Mount(
-		"sysfs",
-		"/sys",
-		"sysfs",
-		0,
-		"",
-	))
-
-	// after top level has been set up
-	mustSyscall("remount root", Mount(
-		"",
-		"/",
-		"",
-		MS_REMOUNT|MS_BIND|
-			MS_RDONLY|MS_NODEV|MS_NOSUID|MS_NOEXEC,
-		"",
-	))
-
-	must(os.WriteFile(
-		"/sys/module/firmware_class/parameters/path",
-		[]byte("/system/lib/firmware"),
-		0,
-	))
-
-}
-
-// mustSyscall calls [log.Fatalln] if err is non-nil.
-func mustSyscall(action string, err error) {
-	if err != nil {
-		log.Fatalln("cannot "+action+":", err)
-	}
-}
-
-// must calls [log.Fatal] with err if it is non-nil.
-func must(err error) {
-	if err != nil {
-		log.Fatal(err)
-	}
-}

cmd/hakurei/command.go

@@ -2,7 +2,6 @@ package main
 
 import (
 	"context"
-	"errors"
 	"fmt"
 	"io"
 	"log"
@@ -12,11 +11,11 @@ import (
 	"strconv"
 	"sync"
 	"time"
+	_ "unsafe" // for go:linkname
 
-	"hakurei.app/check"
 	"hakurei.app/command"
-	"hakurei.app/ext"
+	"hakurei.app/container/check"
-	"hakurei.app/fhs"
+	"hakurei.app/container/fhs"
 	"hakurei.app/hst"
 	"hakurei.app/internal/dbus"
 	"hakurei.app/internal/env"
@@ -27,20 +26,14 @@ import (
 
 // optionalErrorUnwrap calls [errors.Unwrap] and returns the resulting value
 // if it is not nil, or the original value if it is.
-func optionalErrorUnwrap(err error) error {
+//
-	if underlyingErr := errors.Unwrap(err); underlyingErr != nil {
+//go:linkname optionalErrorUnwrap hakurei.app/container.optionalErrorUnwrap
-		return underlyingErr
+func optionalErrorUnwrap(err error) error
-	}
-	return err
-}
-
-var errSuccess = errors.New("success")
 
 func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErrs, out io.Writer) command.Command {
 	var (
-		flagVerbose  bool
+		flagVerbose bool
-		flagInsecure bool
+		flagJSON    bool
-		flagJSON     bool
 	)
 	c := command.New(out, log.Printf, "hakurei", func([]string) error {
 		msg.SwapVerbose(flagVerbose)
@@ -58,7 +51,6 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 		return nil
 	}).
 		Flag(&flagVerbose, "v", command.BoolFlag(false), "Increase log verbosity").
-		Flag(&flagInsecure, "insecure", command.BoolFlag(false), "Allow use of insecure compatibility options").
 		Flag(&flagJSON, "json", command.BoolFlag(false), "Serialise output in JSON when applicable")
 
 	c.Command("shim", command.UsageInternal, func([]string) error { outcome.Shim(msg); return errSuccess })
@@ -67,9 +59,9 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 		var (
 			flagIdentifierFile int
 		)
-		c.NewCommand("run", "Load and start container from configuration file", func(args []string) error {
+		c.NewCommand("app", "Load and start container from configuration file", func(args []string) error {
 			if len(args) < 1 {
-				log.Fatal("run requires at least 1 argument")
+				log.Fatal("app requires at least 1 argument")
 			}
 
 			config := tryPath(msg, args[0])
@@ -77,12 +69,7 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 				config.Container.Args = append(config.Container.Args, args[1:]...)
 			}
 
-			var flags int
+			outcome.Main(ctx, msg, config, flagIdentifierFile)
-			if flagInsecure {
-				flags |= hst.VAllowInsecure
-			}
-
-			outcome.Main(ctx, msg, config, flags, flagIdentifierFile)
 			panic("unreachable")
 		}).
 			Flag(&flagIdentifierFile, "identifier-fd", command.IntFlag(-1),
@@ -102,15 +89,12 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 			flagHomeDir  string
 			flagUserName string
 
-			flagSchedPolicy   string
-			flagSchedPriority int
-
 			flagPrivateRuntime, flagPrivateTmpdir bool
 
 			flagWayland, flagX11, flagDBus, flagPipeWire, flagPulse bool
 		)
 
-		c.NewCommand("exec", "Configure and start a permissive container", func(args []string) error {
+		c.NewCommand("run", "Configure and start a permissive container", func(args []string) error {
 			if flagIdentity < hst.IdentityStart || flagIdentity > hst.IdentityEnd {
 				log.Fatalf("identity %d out of range", flagIdentity)
 			}
@@ -147,12 +131,12 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 					log.Fatal(optionalErrorUnwrap(err))
 					return err
 				} else if progPath, err = check.NewAbs(p); err != nil {
-					log.Fatal(err)
+					log.Fatal(err.Error())
 					return err
 				}
 			}
 
-			var et hst.Enablements
+			var et hst.Enablement
 			if flagWayland {
 				et |= hst.EWayland
 			}
@@ -166,11 +150,11 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 				et |= hst.EPipeWire
 			}
 
-			config := hst.Config{
+			config := &hst.Config{
 				ID:          flagID,
 				Identity:    flagIdentity,
 				Groups:      flagGroups,
-				Enablements: &et,
+				Enablements: hst.NewEnablements(et),
 
 				Container: &hst.ContainerConfig{
 					Filesystem: []hst.FilesystemConfigJSON{
@@ -193,13 +177,6 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 				},
 			}
 
-			if err := config.SchedPolicy.UnmarshalText(
-				[]byte(flagSchedPolicy),
-			); err != nil {
-				log.Fatal(err)
-			}
-			config.SchedPriority = ext.Int(flagSchedPriority)
-
 			// bind GPU stuff
 			if et&(hst.EX11|hst.EWayland) != 0 {
 				config.Container.Filesystem = append(config.Container.Filesystem, hst.FilesystemConfigJSON{FilesystemConfig: &hst.FSBind{
@@ -237,7 +214,7 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 					homeDir = passwd.HomeDir
 				}
 				if a, err := check.NewAbs(homeDir); err != nil {
-					log.Fatal(err)
+					log.Fatal(err.Error())
 					return err
 				} else {
 					config.Container.Home = a
@@ -257,11 +234,11 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 					config.SessionBus = dbus.NewConfig(flagID, true, flagDBusMpris)
 				} else {
 					if f, err := os.Open(flagDBusConfigSession); err != nil {
-						log.Fatal(err)
+						log.Fatal(err.Error())
 					} else {
 						decodeJSON(log.Fatal, "load session bus proxy config", f, &config.SessionBus)
 						if err = f.Close(); err != nil {
-							log.Fatal(err)
+							log.Fatal(err.Error())
 						}
 					}
 				}
@@ -269,11 +246,11 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 				// system bus proxy is optional
 				if flagDBusConfigSystem != "nil" {
 					if f, err := os.Open(flagDBusConfigSystem); err != nil {
-						log.Fatal(err)
+						log.Fatal(err.Error())
 					} else {
 						decodeJSON(log.Fatal, "load system bus proxy config", f, &config.SystemBus)
 						if err = f.Close(); err != nil {
-							log.Fatal(err)
+							log.Fatal(err.Error())
 						}
 					}
 				}
@@ -289,7 +266,7 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 				}
 			}
 
-			outcome.Main(ctx, msg, &config, 0, -1)
+			outcome.Main(ctx, msg, config, -1)
 			panic("unreachable")
 		}).
 			Flag(&flagDBusConfigSession, "dbus-config", command.StringFlag("builtin"),
@@ -310,10 +287,6 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 				"Container home directory").
 			Flag(&flagUserName, "u", command.StringFlag("chronos"),
 				"Passwd user name within sandbox").
-			Flag(&flagSchedPolicy, "policy", command.StringFlag(""),
-				"Scheduling policy to set for the container").
-			Flag(&flagSchedPriority, "priority", command.IntFlag(0),
-				"Scheduling priority to set for the container").
 			Flag(&flagPrivateRuntime, "private-runtime", command.BoolFlag(false),
 				"Do not share XDG_RUNTIME_DIR between containers under the same identity").
 			Flag(&flagPrivateTmpdir, "private-tmpdir", command.BoolFlag(false),
@@ -335,7 +308,7 @@ func buildCommand(ctx context.Context, msg message.Msg, early *earlyHardeningErr
 			flagShort   bool
 			flagNoStore bool
 		)
-		c.NewCommand("show", "Show live or local instance configuration", func(args []string) error {
+		c.NewCommand("show", "Show live or local app configuration", func(args []string) error {
 			switch len(args) {
 			case 0: // system
 				printShowSystem(os.Stdout, flagShort, flagJSON)

cmd/hakurei/command_test.go

@@ -20,12 +20,12 @@ func TestHelp(t *testing.T) {
 	}{
 		{
 			"main", []string{}, `
-Usage:	hakurei [-h | --help] [-v] [--insecure] [--json] COMMAND [OPTIONS]
+Usage:	hakurei [-h | --help] [-v] [--json] COMMAND [OPTIONS]
 
 Commands:
-    run         Load and start container from configuration file
+    app         Load and start container from configuration file
-    exec        Configure and start a permissive container
+    run         Configure and start a permissive container
-    show        Show live or local instance configuration
+    show        Show live or local app configuration
     ps          List active instances
     version     Display version information
     license     Show full license text
@@ -35,8 +35,8 @@ Commands:
 `,
 		},
 		{
-			"exec", []string{"exec", "-h"}, `
+			"run", []string{"run", "-h"}, `
-Usage:	hakurei exec [-h | --help] [--dbus-config <value>] [--dbus-system <value>] [--mpris] [--dbus-log] [--id <value>] [-a <int>] [-g <value>] [-d <value>] [-u <value>] [--policy <value>] [--priority <int>] [--private-runtime] [--private-tmpdir] [--wayland] [-X] [--dbus] [--pipewire] [--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
@@ -60,10 +60,6 @@ Flags:
     	Allow owning MPRIS D-Bus path, has no effect if custom config is available
   -pipewire
     	Enable connection to PipeWire via SecurityContext
-  -policy string
-    	Scheduling policy to set for the container
-  -priority int
-    	Scheduling priority to set for the container
   -private-runtime
     	Do not share XDG_RUNTIME_DIR between containers under the same identity
   -private-tmpdir

cmd/hakurei/json_test.go

@@ -5,7 +5,7 @@ import (
 	"strings"
 	"testing"
 
-	"hakurei.app/internal/stub"
+	"hakurei.app/container/stub"
 )
 
 func TestDecodeJSON(t *testing.T) {

cmd/hakurei/main.go

@@ -1,42 +1,8 @@
-// Hakurei runs user-specified containers as subordinate users.
-//
-// This program is generally invoked by another, higher level program, which
-// creates container configuration via package [hst] or an implementation of it.
-//
-// The parent may leave files open and specify their file descriptor for various
-// uses. In these cases, standard streams and netpoll files are treated as
-// invalid file descriptors and rejected. All string representations must be in
-// decimal.
-//
-// When specifying a [hst.Config] JSON stream or file to the run subcommand, the
-// argument "-" is equivalent to stdin. Otherwise, file descriptor rules
-// described above applies. Invalid file descriptors are treated as file names
-// in their string representation, with the exception that if a netpoll file
-// descriptor is attempted, the program fails.
-//
-// The flag --identifier-fd can be optionally specified to the run subcommand to
-// receive the identifier of the newly started instance. File descriptor rules
-// described above applies, and the file must be writable. This is sent after
-// its state is made available, so the client must not attempt to poll for it.
-// This uses the internal binary format of [hst.ID].
-//
-// For the show and ps subcommands, the flag --json can be applied to the main
-// hakurei command to serialise output in JSON when applicable. Additionally,
-// the flag --short targeting each subcommand is used to omit some information
-// in both JSON and user-facing output. Only JSON-encoded output is covered
-// under the compatibility promise.
-//
-// A template for [hst.Config] demonstrating all available configuration fields
-// is returned by [hst.Template]. The JSON-encoded equivalent of this can be
-// obtained via the template subcommand. Fields left unpopulated in the template
-// (the direct_* family of fields, which are insecure under any configuration if
-// enabled) are unsupported.
-//
-// For simple (but insecure) testing scenarios, the exec subcommand can be used
-// to generate a simple, permissive configuration in-memory. See its help
-// message for all available options.
 package main
 
+// this works around go:embed '..' limitation
+//go:generate cp ../../LICENSE .
+
 import (
 	"context"
 	_ "embed"
@@ -47,13 +13,15 @@ import (
 	"syscall"
 
 	"hakurei.app/container"
-	"hakurei.app/ext"
 	"hakurei.app/message"
 )
 
-//go:generate cp ../../LICENSE .
+var (
-//go:embed LICENSE
+	errSuccess = errors.New("success")
-var license string
+
+	//go:embed LICENSE
+	license string
+)
 
 // earlyHardeningErrs are errors collected while setting up early hardening feature.
 type earlyHardeningErrs struct{ yamaLSM, dumpable error }
@@ -62,13 +30,13 @@ func main() {
 	// early init path, skips root check and duplicate PR_SET_DUMPABLE
 	container.TryArgv0(nil)
 
-	log.SetFlags(0)
 	log.SetPrefix("hakurei: ")
+	log.SetFlags(0)
 	msg := message.New(log.Default())
 
 	early := earlyHardeningErrs{
-		yamaLSM:  ext.SetPtracer(0),
+		yamaLSM:  container.SetPtracer(0),
-		dumpable: ext.SetDumpable(ext.SUID_DUMP_DISABLE),
+		dumpable: container.SetDumpable(container.SUID_DUMP_DISABLE),
 	}
 
 	if os.Geteuid() == 0 {

cmd/hakurei/parse.go

@@ -17,9 +17,8 @@ import (
 )
 
 // tryPath attempts to read [hst.Config] from multiple sources.
-//
+// tryPath reads from [os.Stdin] if name has value "-".
-// tryPath reads from [os.Stdin] if name has value "-". Otherwise, name is
+// Otherwise, name is passed to tryFd, and if that returns nil, name is passed to [os.Open].
-// passed to tryFd, and if that returns nil, name is passed to [os.Open].
 func tryPath(msg message.Msg, name string) (config *hst.Config) {
 	var r io.ReadCloser
 	config = new(hst.Config)
@@ -47,8 +46,7 @@ func tryPath(msg message.Msg, name string) (config *hst.Config) {
 	return
 }
 
-// tryFd returns a [io.ReadCloser] if name represents an integer corresponding
+// tryFd returns a [io.ReadCloser] if name represents an integer corresponding to a valid file descriptor.
-// to a valid file descriptor.
 func tryFd(msg message.Msg, name string) io.ReadCloser {
 	if v, err := strconv.Atoi(name); err != nil {
 		if !errors.Is(err, strconv.ErrSyntax) {
@@ -62,12 +60,7 @@ func tryFd(msg message.Msg, name string) io.ReadCloser {
 
 		msg.Verbosef("trying config stream from %d", v)
 		fd := uintptr(v)
-		if _, _, errno := syscall.Syscall(
+		if _, _, errno := syscall.Syscall(syscall.SYS_FCNTL, fd, syscall.F_GETFD, 0); errno != 0 {
-			syscall.SYS_FCNTL,
-			fd,
-			syscall.F_GETFD,
-			0,
-		); errno != 0 {
 			if errors.Is(errno, syscall.EBADF) { // reject bad fd
 				return nil
 			}
@@ -82,12 +75,10 @@ func tryFd(msg message.Msg, name string) io.ReadCloser {
 	}
 }
 
-// shortLengthMin is the minimum length a short form identifier can have and
+// shortLengthMin is the minimum length a short form identifier can have and still be interpreted as an identifier.
-// still be interpreted as an identifier.
 const shortLengthMin = 1 << 3
 
-// shortIdentifier returns an eight character short representation of [hst.ID]
+// shortIdentifier returns an eight character short representation of [hst.ID] from its random bytes.
-// from its random bytes.
 func shortIdentifier(id *hst.ID) string {
 	return shortIdentifierString(id.String())
 }
@@ -97,8 +88,7 @@ func shortIdentifierString(s string) string {
 	return s[len(hst.ID{}) : len(hst.ID{})+shortLengthMin]
 }
 
-// tryIdentifier attempts to match [hst.State] from a [hex] representation of
+// tryIdentifier attempts to match [hst.State] from a [hex] representation of [hst.ID] or a prefix of its lower half.
-// [hst.ID] or a prefix of its lower half.
 func tryIdentifier(msg message.Msg, name string, s *store.Store) *hst.State {
 	const (
 		likeShort = 1 << iota
@@ -106,8 +96,7 @@ func tryIdentifier(msg message.Msg, name string, s *store.Store) *hst.State {
 	)
 
 	var likely uintptr
-	// half the hex representation
+	if len(name) >= shortLengthMin && len(name) <= len(hst.ID{}) { // half the hex representation
-	if len(name) >= shortLengthMin && len(name) <= len(hst.ID{}) {
 		// cannot safely decode here due to unknown alignment
 		for _, c := range name {
 			if c >= '0' && c <= '9' {

cmd/hakurei/parse_test.go

@@ -6,7 +6,7 @@ import (
 	"testing"
 	"time"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
 	"hakurei.app/hst"
 	"hakurei.app/internal/store"
 	"hakurei.app/message"

cmd/hakurei/print.go

@@ -56,7 +56,7 @@ func printShowInstance(
 	t := newPrinter(output)
 	defer t.MustFlush()
 
-	if err := config.Validate(hst.VAllowInsecure); err != nil {
+	if err := config.Validate(); err != nil {
 		valid = false
 		if m, ok := message.GetMessage(err); ok {
 			mustPrint(output, "Error: "+m+"!\n\n")

cmd/hakurei/print_test.go

@@ -7,7 +7,7 @@ import (
 	"testing"
 	"time"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
 	"hakurei.app/hst"
 	"hakurei.app/internal/store"
 	"hakurei.app/message"
@@ -32,7 +32,7 @@ var (
 		PID:     0xbeef,
 		ShimPID: 0xcafe,
 		Config: &hst.Config{
-			Enablements: new(hst.EWayland | hst.EPipeWire),
+			Enablements: hst.NewEnablements(hst.EWayland | hst.EPipeWire),
 			Identity:    1,
 			Container: &hst.ContainerConfig{
 				Shell: check.MustAbs("/bin/sh"),

cmd/hpkg/README

@@ -0,0 +1,7 @@
+This program is a proof of concept and is now deprecated. It is only kept
+around for API demonstration purposes and to make the most out of the test
+suite.
+
+This program is replaced by planterette, which can be found at
+https://git.gensokyo.uk/security/planterette. Development effort should be
+focused there instead.

cmd/hpkg/app.go

@@ -0,0 +1,173 @@
+package main
+
+import (
+	"encoding/json"
+	"log"
+	"os"
+
+	"hakurei.app/container/check"
+	"hakurei.app/container/fhs"
+	"hakurei.app/hst"
+)
+
+type appInfo struct {
+	Name    string `json:"name"`
+	Version string `json:"version"`
+
+	// passed through to [hst.Config]
+	ID string `json:"id"`
+	// passed through to [hst.Config]
+	Identity int `json:"identity"`
+	// passed through to [hst.Config]
+	Groups []string `json:"groups,omitempty"`
+	// passed through to [hst.Config]
+	Devel bool `json:"devel,omitempty"`
+	// passed through to [hst.Config]
+	Userns bool `json:"userns,omitempty"`
+	// passed through to [hst.Config]
+	HostNet bool `json:"net,omitempty"`
+	// passed through to [hst.Config]
+	HostAbstract bool `json:"abstract,omitempty"`
+	// passed through to [hst.Config]
+	Device bool `json:"dev,omitempty"`
+	// passed through to [hst.Config]
+	Tty bool `json:"tty,omitempty"`
+	// passed through to [hst.Config]
+	MapRealUID bool `json:"map_real_uid,omitempty"`
+	// passed through to [hst.Config]
+	DirectWayland bool `json:"direct_wayland,omitempty"`
+	// passed through to [hst.Config]
+	SystemBus *hst.BusConfig `json:"system_bus,omitempty"`
+	// passed through to [hst.Config]
+	SessionBus *hst.BusConfig `json:"session_bus,omitempty"`
+	// passed through to [hst.Config]
+	Enablements *hst.Enablements `json:"enablements,omitempty"`
+
+	// passed through to [hst.Config]
+	Multiarch bool `json:"multiarch,omitempty"`
+	// passed through to [hst.Config]
+	Bluetooth bool `json:"bluetooth,omitempty"`
+
+	// allow gpu access within sandbox
+	GPU bool `json:"gpu"`
+	// store path to nixGL mesa wrappers
+	Mesa string `json:"mesa,omitempty"`
+	// store path to nixGL source
+	NixGL string `json:"nix_gl,omitempty"`
+	// store path to activate-and-exec script
+	Launcher *check.Absolute `json:"launcher"`
+	// store path to /run/current-system
+	CurrentSystem *check.Absolute `json:"current_system"`
+	// store path to home-manager activation package
+	ActivationPackage string `json:"activation_package"`
+}
+
+func (app *appInfo) toHst(pathSet *appPathSet, pathname *check.Absolute, argv []string, flagDropShell bool) *hst.Config {
+	config := &hst.Config{
+		ID: app.ID,
+
+		Enablements: app.Enablements,
+
+		SystemBus:     app.SystemBus,
+		SessionBus:    app.SessionBus,
+		DirectWayland: app.DirectWayland,
+
+		Identity: app.Identity,
+		Groups:   app.Groups,
+
+		Container: &hst.ContainerConfig{
+			Hostname: formatHostname(app.Name),
+			Filesystem: []hst.FilesystemConfigJSON{
+				{FilesystemConfig: &hst.FSBind{Target: fhs.AbsEtc, Source: pathSet.cacheDir.Append("etc"), Special: true}},
+				{FilesystemConfig: &hst.FSBind{Source: pathSet.nixPath.Append("store"), Target: pathNixStore}},
+				{FilesystemConfig: &hst.FSLink{Target: pathCurrentSystem, Linkname: app.CurrentSystem.String()}},
+				{FilesystemConfig: &hst.FSLink{Target: pathBin, Linkname: pathSwBin.String()}},
+				{FilesystemConfig: &hst.FSLink{Target: fhs.AbsUsrBin, Linkname: pathSwBin.String()}},
+				{FilesystemConfig: &hst.FSBind{Source: pathSet.metaPath, Target: hst.AbsPrivateTmp.Append("app")}},
+				{FilesystemConfig: &hst.FSBind{Source: fhs.AbsEtc.Append("resolv.conf"), Optional: true}},
+				{FilesystemConfig: &hst.FSBind{Source: fhs.AbsSys.Append("block"), Optional: true}},
+				{FilesystemConfig: &hst.FSBind{Source: fhs.AbsSys.Append("bus"), Optional: true}},
+				{FilesystemConfig: &hst.FSBind{Source: fhs.AbsSys.Append("class"), Optional: true}},
+				{FilesystemConfig: &hst.FSBind{Source: fhs.AbsSys.Append("dev"), Optional: true}},
+				{FilesystemConfig: &hst.FSBind{Source: fhs.AbsSys.Append("devices"), Optional: true}},
+				{FilesystemConfig: &hst.FSBind{Target: pathDataData.Append(app.ID), Source: pathSet.homeDir, Write: true, Ensure: true}},
+			},
+
+			Username: "hakurei",
+			Shell:    pathShell,
+			Home:     pathDataData.Append(app.ID),
+
+			Path: pathname,
+			Args: argv,
+		},
+		ExtraPerms: []hst.ExtraPermConfig{
+			{Path: dataHome, Execute: true},
+			{Ensure: true, Path: pathSet.baseDir, Read: true, Write: true, Execute: true},
+		},
+	}
+
+	if app.Devel {
+		config.Container.Flags |= hst.FDevel
+	}
+	if app.Userns {
+		config.Container.Flags |= hst.FUserns
+	}
+	if app.HostNet {
+		config.Container.Flags |= hst.FHostNet
+	}
+	if app.HostAbstract {
+		config.Container.Flags |= hst.FHostAbstract
+	}
+	if app.Device {
+		config.Container.Flags |= hst.FDevice
+	}
+	if app.Tty || flagDropShell {
+		config.Container.Flags |= hst.FTty
+	}
+	if app.MapRealUID {
+		config.Container.Flags |= hst.FMapRealUID
+	}
+	if app.Multiarch {
+		config.Container.Flags |= hst.FMultiarch
+	}
+	config.Container.Flags |= hst.FShareRuntime | hst.FShareTmpdir
+	return config
+}
+
+func loadAppInfo(name string, beforeFail func()) *appInfo {
+	bundle := new(appInfo)
+	if f, err := os.Open(name); err != nil {
+		beforeFail()
+		log.Fatalf("cannot open bundle: %v", err)
+	} else if err = json.NewDecoder(f).Decode(&bundle); err != nil {
+		beforeFail()
+		log.Fatalf("cannot parse bundle metadata: %v", err)
+	} else if err = f.Close(); err != nil {
+		log.Printf("cannot close bundle metadata: %v", err)
+		// not fatal
+	}
+
+	if bundle.ID == "" {
+		beforeFail()
+		log.Fatal("application identifier must not be empty")
+	}
+	if bundle.Launcher == nil {
+		beforeFail()
+		log.Fatal("launcher must not be empty")
+	}
+	if bundle.CurrentSystem == nil {
+		beforeFail()
+		log.Fatal("current-system must not be empty")
+	}
+
+	return bundle
+}
+
+func formatHostname(name string) string {
+	if h, err := os.Hostname(); err != nil {
+		log.Printf("cannot get hostname: %v", err)
+		return "hakurei-" + name
+	} else {
+		return h + "-" + name
+	}
+}

cmd/hpkg/build.nix

@@ -0,0 +1,256 @@
+{
+  nixpkgsFor,
+  system,
+  nixpkgs,
+  home-manager,
+}:
+
+{
+  lib,
+  stdenv,
+  closureInfo,
+  writeScript,
+  runtimeShell,
+  writeText,
+  symlinkJoin,
+  vmTools,
+  runCommand,
+  fetchFromGitHub,
+
+  zstd,
+  nix,
+  sqlite,
+
+  name ? throw "name is required",
+  version ? throw "version is required",
+  pname ? "${name}-${version}",
+  modules ? [ ],
+  nixosModules ? [ ],
+  script ? ''
+    exec "$SHELL" "$@"
+  '',
+
+  id ? name,
+  identity ? throw "identity is required",
+  groups ? [ ],
+  userns ? false,
+  net ? true,
+  dev ? false,
+  no_new_session ? false,
+  map_real_uid ? false,
+  direct_wayland ? false,
+  system_bus ? null,
+  session_bus ? null,
+
+  allow_wayland ? true,
+  allow_x11 ? false,
+  allow_dbus ? true,
+  allow_audio ? true,
+  gpu ? allow_wayland || allow_x11,
+}:
+
+let
+  inherit (lib) optionals;
+
+  homeManagerConfiguration = home-manager.lib.homeManagerConfiguration {
+    pkgs = nixpkgsFor.${system};
+    modules = modules ++ [
+      {
+        home = {
+          username = "hakurei";
+          homeDirectory = "/data/data/${id}";
+          stateVersion = "22.11";
+        };
+      }
+    ];
+  };
+
+  launcher = writeScript "hakurei-${pname}" ''
+    #!${runtimeShell} -el
+    ${script}
+  '';
+
+  extraNixOSConfig =
+    { pkgs, ... }:
+    {
+      environment = {
+        etc.nixpkgs.source = nixpkgs.outPath;
+        systemPackages = [ pkgs.nix ];
+      };
+
+      imports = nixosModules;
+    };
+  nixos = nixpkgs.lib.nixosSystem {
+    inherit system;
+    modules = [
+      extraNixOSConfig
+      { nix.settings.experimental-features = [ "flakes" ]; }
+      { nix.settings.experimental-features = [ "nix-command" ]; }
+      { boot.isContainer = true; }
+      { system.stateVersion = "22.11"; }
+    ];
+  };
+
+  etc = vmTools.runInLinuxVM (
+    runCommand "etc" { } ''
+      mkdir -p /etc
+      ${nixos.config.system.build.etcActivationCommands}
+
+      # remove unused files
+      rm -rf /etc/sudoers
+
+      mkdir -p $out
+      tar -C /etc -cf "$out/etc.tar" .
+    ''
+  );
+
+  extendSessionDefault = id: ext: {
+    filter = true;
+
+    talk = [ "org.freedesktop.Notifications" ] ++ ext.talk;
+    own =
+      (optionals (id != null) [
+        "${id}.*"
+        "org.mpris.MediaPlayer2.${id}.*"
+      ])
+      ++ ext.own;
+
+    inherit (ext) call broadcast;
+  };
+
+  nixGL = fetchFromGitHub {
+    owner = "nix-community";
+    repo = "nixGL";
+    rev = "310f8e49a149e4c9ea52f1adf70cdc768ec53f8a";
+    hash = "sha256-lnzZQYG0+EXl/6NkGpyIz+FEOc/DSEG57AP1VsdeNrM=";
+  };
+
+  mesaWrappers =
+    let
+      isIntelX86Platform = system == "x86_64-linux";
+      nixGLPackages = import (nixGL + "/default.nix") {
+        pkgs = nixpkgs.legacyPackages.${system};
+        enable32bits = isIntelX86Platform;
+        enableIntelX86Extensions = isIntelX86Platform;
+      };
+    in
+    symlinkJoin {
+      name = "nixGL-mesa";
+      paths = with nixGLPackages; [
+        nixGLIntel
+        nixVulkanIntel
+      ];
+    };
+
+  info = builtins.toJSON {
+    inherit
+      name
+      version
+      id
+      identity
+      launcher
+      groups
+      userns
+      net
+      dev
+      no_new_session
+      map_real_uid
+      direct_wayland
+      system_bus
+      gpu
+      ;
+
+    session_bus =
+      if session_bus != null then
+        (session_bus (extendSessionDefault id))
+      else
+        (extendSessionDefault id {
+          talk = [ ];
+          own = [ ];
+          call = { };
+          broadcast = { };
+        });
+
+    enablements = {
+      wayland = allow_wayland;
+      x11 = allow_x11;
+      dbus = allow_dbus;
+      pipewire = allow_audio;
+    };
+
+    mesa = if gpu then mesaWrappers else null;
+    nix_gl = if gpu then nixGL else null;
+    current_system = nixos.config.system.build.toplevel;
+    activation_package = homeManagerConfiguration.activationPackage;
+  };
+in
+
+stdenv.mkDerivation {
+  name = "${pname}.pkg";
+  inherit version;
+  __structuredAttrs = true;
+
+  nativeBuildInputs = [
+    zstd
+    nix
+    sqlite
+  ];
+
+  buildCommand = ''
+    NIX_ROOT="$(mktemp -d)"
+    export USER="nobody"
+
+    # create bootstrap store
+    bootstrapClosureInfo="${
+      closureInfo {
+        rootPaths = [
+          nix
+          nixos.config.system.build.toplevel
+        ];
+      }
+    }"
+    echo "copying bootstrap store paths..."
+    mkdir -p "$NIX_ROOT/nix/store"
+    xargs -n 1 -a "$bootstrapClosureInfo/store-paths" cp -at "$NIX_ROOT/nix/store/"
+    NIX_REMOTE="local?root=$NIX_ROOT" nix-store --load-db < "$bootstrapClosureInfo/registration"
+    NIX_REMOTE="local?root=$NIX_ROOT" nix-store --optimise
+    sqlite3 "$NIX_ROOT/nix/var/nix/db/db.sqlite" "UPDATE ValidPaths SET registrationTime = ''${SOURCE_DATE_EPOCH}"
+    chmod -R +r "$NIX_ROOT/nix/var"
+
+    # create binary cache
+    closureInfo="${
+      closureInfo {
+        rootPaths = [
+          homeManagerConfiguration.activationPackage
+          launcher
+        ]
+        ++ optionals gpu [
+          mesaWrappers
+          nixGL
+        ];
+      }
+    }"
+    echo "copying application paths..."
+    TMP_STORE="$(mktemp -d)"
+    mkdir -p "$TMP_STORE/nix/store"
+    xargs -n 1 -a "$closureInfo/store-paths" cp -at "$TMP_STORE/nix/store/"
+    NIX_REMOTE="local?root=$TMP_STORE" nix-store --load-db < "$closureInfo/registration"
+    sqlite3 "$TMP_STORE/nix/var/nix/db/db.sqlite" "UPDATE ValidPaths SET registrationTime = ''${SOURCE_DATE_EPOCH}"
+    NIX_REMOTE="local?root=$TMP_STORE" nix --offline --extra-experimental-features nix-command \
+        --verbose --log-format raw-with-logs \
+        copy --all --no-check-sigs --to \
+        "file://$NIX_ROOT/res?compression=zstd&compression-level=19&parallel-compression=true"
+
+    # package /etc
+    mkdir -p "$NIX_ROOT/etc"
+    tar -C "$NIX_ROOT/etc" -xf "${etc}/etc.tar"
+
+    # write metadata
+    cp "${writeText "bundle.json" info}" "$NIX_ROOT/bundle.json"
+
+    # create an intermediate file to improve zstd performance
+    INTER="$(mktemp)"
+    tar -C "$NIX_ROOT" -cf "$INTER" .
+    zstd -T0 -19 -fo "$out" "$INTER"
+  '';
+}

cmd/hpkg/main.go

@@ -0,0 +1,335 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"log"
+	"os"
+	"os/signal"
+	"path"
+	"syscall"
+
+	"hakurei.app/command"
+	"hakurei.app/container/check"
+	"hakurei.app/container/fhs"
+	"hakurei.app/hst"
+	"hakurei.app/message"
+)
+
+var (
+	errSuccess = errors.New("success")
+)
+
+func main() {
+	log.SetPrefix("hpkg: ")
+	log.SetFlags(0)
+	msg := message.New(log.Default())
+
+	if err := os.Setenv("SHELL", pathShell.String()); err != nil {
+		log.Fatalf("cannot set $SHELL: %v", err)
+	}
+
+	if os.Geteuid() == 0 {
+		log.Fatal("this program must not run as root")
+	}
+
+	ctx, stop := signal.NotifyContext(context.Background(),
+		syscall.SIGINT, syscall.SIGTERM)
+	defer stop() // unreachable
+
+	var (
+		flagVerbose   bool
+		flagDropShell bool
+	)
+	c := command.New(os.Stderr, log.Printf, "hpkg", func([]string) error { msg.SwapVerbose(flagVerbose); return nil }).
+		Flag(&flagVerbose, "v", command.BoolFlag(false), "Print debug messages to the console").
+		Flag(&flagDropShell, "s", command.BoolFlag(false), "Drop to a shell in place of next hakurei action")
+
+	{
+		var (
+			flagDropShellActivate bool
+		)
+		c.NewCommand("install", "Install an application from its package", func(args []string) error {
+			if len(args) != 1 {
+				log.Println("invalid argument")
+				return syscall.EINVAL
+			}
+			pkgPath := args[0]
+			if !path.IsAbs(pkgPath) {
+				if dir, err := os.Getwd(); err != nil {
+					log.Printf("cannot get current directory: %v", err)
+					return err
+				} else {
+					pkgPath = path.Join(dir, pkgPath)
+				}
+			}
+
+			/*
+				Look up paths to programs started by hpkg.
+				This is done here to ease error handling as cleanup is not yet required.
+			*/
+
+			var (
+				_     = lookPath("zstd")
+				tar   = lookPath("tar")
+				chmod = lookPath("chmod")
+				rm    = lookPath("rm")
+			)
+
+			/*
+				Extract package and set up for cleanup.
+			*/
+
+			var workDir *check.Absolute
+			if p, err := os.MkdirTemp("", "hpkg.*"); err != nil {
+				log.Printf("cannot create temporary directory: %v", err)
+				return err
+			} else if workDir, err = check.NewAbs(p); err != nil {
+				log.Printf("invalid temporary directory: %v", err)
+				return err
+			}
+			cleanup := func() {
+				// should be faster than a native implementation
+				mustRun(msg, chmod, "-R", "+w", workDir.String())
+				mustRun(msg, rm, "-rf", workDir.String())
+			}
+			beforeRunFail.Store(&cleanup)
+
+			mustRun(msg, tar, "-C", workDir.String(), "-xf", pkgPath)
+
+			/*
+				Parse bundle and app metadata, do pre-install checks.
+			*/
+
+			bundle := loadAppInfo(path.Join(workDir.String(), "bundle.json"), cleanup)
+			pathSet := pathSetByApp(bundle.ID)
+
+			a := bundle
+			if s, err := os.Stat(pathSet.metaPath.String()); err != nil {
+				if !os.IsNotExist(err) {
+					cleanup()
+					log.Printf("cannot access %q: %v", pathSet.metaPath, err)
+					return err
+				}
+				// did not modify app, clean installation condition met later
+			} else if s.IsDir() {
+				cleanup()
+				log.Printf("metadata path %q is not a file", pathSet.metaPath)
+				return syscall.EBADMSG
+			} else {
+				a = loadAppInfo(pathSet.metaPath.String(), cleanup)
+				if a.ID != bundle.ID {
+					cleanup()
+					log.Printf("app %q claims to have identifier %q",
+						bundle.ID, a.ID)
+					return syscall.EBADE
+				}
+				// sec: should verify credentials
+			}
+
+			if a != bundle {
+				// do not try to re-install
+				if a.NixGL == bundle.NixGL &&
+					a.CurrentSystem == bundle.CurrentSystem &&
+					a.Launcher == bundle.Launcher &&
+					a.ActivationPackage == bundle.ActivationPackage {
+					cleanup()
+					log.Printf("package %q is identical to local application %q",
+						pkgPath, a.ID)
+					return errSuccess
+				}
+
+				// identity determines uid
+				if a.Identity != bundle.Identity {
+					cleanup()
+					log.Printf("package %q identity %d differs from installed %d",
+						pkgPath, bundle.Identity, a.Identity)
+					return syscall.EBADE
+				}
+
+				// sec: should compare version string
+				msg.Verbosef("installing application %q version %q over local %q",
+					bundle.ID, bundle.Version, a.Version)
+			} else {
+				msg.Verbosef("application %q clean installation", bundle.ID)
+				// sec: should install credentials
+			}
+
+			/*
+				Setup steps for files owned by the target user.
+			*/
+
+			withCacheDir(ctx, msg, "install", []string{
+				// export inner bundle path in the environment
+				"export BUNDLE=" + hst.PrivateTmp + "/bundle",
+				// replace inner /etc
+				"mkdir -p etc",
+				"chmod -R +w etc",
+				"rm -rf etc",
+				"cp -dRf $BUNDLE/etc etc",
+				// replace inner /nix
+				"mkdir -p nix",
+				"chmod -R +w nix",
+				"rm -rf nix",
+				"cp -dRf /nix nix",
+				// copy from binary cache
+				"nix copy --offline --no-check-sigs --all --from file://$BUNDLE/res --to $PWD",
+				// deduplicate nix store
+				"nix store --offline --store $PWD optimise",
+				// make cache directory world-readable for autoetc
+				"chmod 0755 .",
+			}, workDir, bundle, pathSet, flagDropShell, cleanup)
+
+			if bundle.GPU {
+				withCacheDir(ctx, msg, "mesa-wrappers", []string{
+					// link nixGL mesa wrappers
+					"mkdir -p nix/.nixGL",
+					"ln -s " + bundle.Mesa + "/bin/nixGLIntel nix/.nixGL/nixGL",
+					"ln -s " + bundle.Mesa + "/bin/nixVulkanIntel nix/.nixGL/nixVulkan",
+				}, workDir, bundle, pathSet, false, cleanup)
+			}
+
+			/*
+				Activate home-manager generation.
+			*/
+
+			withNixDaemon(ctx, msg, "activate", []string{
+				// clean up broken links
+				"mkdir -p .local/state/{nix,home-manager}",
+				"chmod -R +w .local/state/{nix,home-manager}",
+				"rm -rf .local/state/{nix,home-manager}",
+				// run activation script
+				bundle.ActivationPackage + "/activate",
+			}, false, func(config *hst.Config) *hst.Config { return config },
+				bundle, pathSet, flagDropShellActivate, cleanup)
+
+			/*
+				Installation complete. Write metadata to block re-installs or downgrades.
+			*/
+
+			// serialise metadata to ensure consistency
+			if f, err := os.OpenFile(pathSet.metaPath.String()+"~", os.O_CREATE|os.O_WRONLY|os.O_TRUNC, 0644); err != nil {
+				cleanup()
+				log.Printf("cannot create metadata file: %v", err)
+				return err
+			} else if err = json.NewEncoder(f).Encode(bundle); err != nil {
+				cleanup()
+				log.Printf("cannot write metadata: %v", err)
+				return err
+			} else if err = f.Close(); err != nil {
+				log.Printf("cannot close metadata file: %v", err)
+				// not fatal
+			}
+
+			if err := os.Rename(pathSet.metaPath.String()+"~", pathSet.metaPath.String()); err != nil {
+				cleanup()
+				log.Printf("cannot rename metadata file: %v", err)
+				return err
+			}
+
+			cleanup()
+			return errSuccess
+		}).
+			Flag(&flagDropShellActivate, "s", command.BoolFlag(false), "Drop to a shell on activation")
+	}
+
+	{
+		var (
+			flagDropShellNixGL bool
+			flagAutoDrivers    bool
+		)
+		c.NewCommand("start", "Start an application", func(args []string) error {
+			if len(args) < 1 {
+				log.Println("invalid argument")
+				return syscall.EINVAL
+			}
+
+			/*
+				Parse app metadata.
+			*/
+
+			id := args[0]
+			pathSet := pathSetByApp(id)
+			a := loadAppInfo(pathSet.metaPath.String(), func() {})
+			if a.ID != id {
+				log.Printf("app %q claims to have identifier %q", id, a.ID)
+				return syscall.EBADE
+			}
+
+			/*
+				Prepare nixGL.
+			*/
+
+			if a.GPU && flagAutoDrivers {
+				withNixDaemon(ctx, msg, "nix-gl", []string{
+					"mkdir -p /nix/.nixGL/auto",
+					"rm -rf /nix/.nixGL/auto",
+					"export NIXPKGS_ALLOW_UNFREE=1",
+					"nix build --impure " +
+						"--out-link /nix/.nixGL/auto/opengl " +
+						"--override-input nixpkgs path:/etc/nixpkgs " +
+						"path:" + a.NixGL,
+					"nix build --impure " +
+						"--out-link /nix/.nixGL/auto/vulkan " +
+						"--override-input nixpkgs path:/etc/nixpkgs " +
+						"path:" + a.NixGL + "#nixVulkanNvidia",
+				}, true, func(config *hst.Config) *hst.Config {
+					config.Container.Filesystem = append(config.Container.Filesystem, []hst.FilesystemConfigJSON{
+						{FilesystemConfig: &hst.FSBind{Source: fhs.AbsEtc.Append("resolv.conf"), Optional: true}},
+						{FilesystemConfig: &hst.FSBind{Source: fhs.AbsSys.Append("block"), Optional: true}},
+						{FilesystemConfig: &hst.FSBind{Source: fhs.AbsSys.Append("bus"), Optional: true}},
+						{FilesystemConfig: &hst.FSBind{Source: fhs.AbsSys.Append("class"), Optional: true}},
+						{FilesystemConfig: &hst.FSBind{Source: fhs.AbsSys.Append("dev"), Optional: true}},
+						{FilesystemConfig: &hst.FSBind{Source: fhs.AbsSys.Append("devices"), Optional: true}},
+					}...)
+					appendGPUFilesystem(config)
+					return config
+				}, a, pathSet, flagDropShellNixGL, func() {})
+			}
+
+			/*
+				Create app configuration.
+			*/
+
+			pathname := a.Launcher
+			argv := make([]string, 1, len(args))
+			if flagDropShell {
+				pathname = pathShell
+				argv[0] = bash
+			} else {
+				argv[0] = a.Launcher.String()
+			}
+			argv = append(argv, args[1:]...)
+			config := a.toHst(pathSet, pathname, argv, flagDropShell)
+
+			/*
+				Expose GPU devices.
+			*/
+
+			if a.GPU {
+				config.Container.Filesystem = append(config.Container.Filesystem,
+					hst.FilesystemConfigJSON{FilesystemConfig: &hst.FSBind{Source: pathSet.nixPath.Append(".nixGL"), Target: hst.AbsPrivateTmp.Append("nixGL")}})
+				appendGPUFilesystem(config)
+			}
+
+			/*
+				Spawn app.
+			*/
+
+			mustRunApp(ctx, msg, config, func() {})
+			return errSuccess
+		}).
+			Flag(&flagDropShellNixGL, "s", command.BoolFlag(false), "Drop to a shell on nixGL build").
+			Flag(&flagAutoDrivers, "auto-drivers", command.BoolFlag(false), "Attempt automatic opengl driver detection")
+	}
+
+	c.MustParse(os.Args[1:], func(err error) {
+		msg.Verbosef("command returned %v", err)
+		if errors.Is(err, errSuccess) {
+			msg.BeforeExit()
+			os.Exit(0)
+		}
+	})
+	log.Fatal("unreachable")
+}

cmd/hpkg/paths.go

@@ -0,0 +1,117 @@
+package main
+
+import (
+	"log"
+	"os"
+	"os/exec"
+	"strconv"
+	"sync/atomic"
+
+	"hakurei.app/container/check"
+	"hakurei.app/container/fhs"
+	"hakurei.app/hst"
+	"hakurei.app/message"
+)
+
+const bash = "bash"
+
+var (
+	dataHome *check.Absolute
+)
+
+func init() {
+	// dataHome
+	if a, err := check.NewAbs(os.Getenv("HAKUREI_DATA_HOME")); err == nil {
+		dataHome = a
+	} else {
+		dataHome = fhs.AbsVarLib.Append("hakurei/" + strconv.Itoa(os.Getuid()))
+	}
+}
+
+var (
+	pathBin = fhs.AbsRoot.Append("bin")
+
+	pathNix           = check.MustAbs("/nix/")
+	pathNixStore      = pathNix.Append("store/")
+	pathCurrentSystem = fhs.AbsRun.Append("current-system")
+	pathSwBin         = pathCurrentSystem.Append("sw/bin/")
+	pathShell         = pathSwBin.Append(bash)
+
+	pathData     = check.MustAbs("/data")
+	pathDataData = pathData.Append("data")
+)
+
+func lookPath(file string) string {
+	if p, err := exec.LookPath(file); err != nil {
+		log.Fatalf("%s: command not found", file)
+		return ""
+	} else {
+		return p
+	}
+}
+
+var beforeRunFail = new(atomic.Pointer[func()])
+
+func mustRun(msg message.Msg, name string, arg ...string) {
+	msg.Verbosef("spawning process: %q %q", name, arg)
+	cmd := exec.Command(name, arg...)
+	cmd.Stdin, cmd.Stdout, cmd.Stderr = os.Stdin, os.Stdout, os.Stderr
+	if err := cmd.Run(); err != nil {
+		if f := beforeRunFail.Swap(nil); f != nil {
+			(*f)()
+		}
+		log.Fatalf("%s: %v", name, err)
+	}
+}
+
+type appPathSet struct {
+	// ${dataHome}/${id}
+	baseDir *check.Absolute
+	// ${baseDir}/app
+	metaPath *check.Absolute
+	// ${baseDir}/files
+	homeDir *check.Absolute
+	// ${baseDir}/cache
+	cacheDir *check.Absolute
+	// ${baseDir}/cache/nix
+	nixPath *check.Absolute
+}
+
+func pathSetByApp(id string) *appPathSet {
+	pathSet := new(appPathSet)
+	pathSet.baseDir = dataHome.Append(id)
+	pathSet.metaPath = pathSet.baseDir.Append("app")
+	pathSet.homeDir = pathSet.baseDir.Append("files")
+	pathSet.cacheDir = pathSet.baseDir.Append("cache")
+	pathSet.nixPath = pathSet.cacheDir.Append("nix")
+	return pathSet
+}
+
+func appendGPUFilesystem(config *hst.Config) {
+	config.Container.Filesystem = append(config.Container.Filesystem, []hst.FilesystemConfigJSON{
+		// flatpak commit 763a686d874dd668f0236f911de00b80766ffe79
+		{FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("dri"), Device: true, Optional: true}},
+		// mali
+		{FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("mali"), Device: true, Optional: true}},
+		{FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("mali0"), Device: true, Optional: true}},
+		{FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("umplock"), Device: true, Optional: true}},
+		// nvidia
+		{FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidiactl"), Device: true, Optional: true}},
+		{FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia-modeset"), Device: true, Optional: true}},
+		// nvidia OpenCL/CUDA
+		{FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia-uvm"), Device: true, Optional: true}},
+		{FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia-uvm-tools"), Device: true, Optional: true}},
+
+		// flatpak commit d2dff2875bb3b7e2cd92d8204088d743fd07f3ff
+		{FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia0"), Device: true, Optional: true}}, {FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia1"), Device: true, Optional: true}},
+		{FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia2"), Device: true, Optional: true}}, {FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia3"), Device: true, Optional: true}},
+		{FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia4"), Device: true, Optional: true}}, {FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia5"), Device: true, Optional: true}},
+		{FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia6"), Device: true, Optional: true}}, {FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia7"), Device: true, Optional: true}},
+		{FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia8"), Device: true, Optional: true}}, {FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia9"), Device: true, Optional: true}},
+		{FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia10"), Device: true, Optional: true}}, {FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia11"), Device: true, Optional: true}},
+		{FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia12"), Device: true, Optional: true}}, {FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia13"), Device: true, Optional: true}},
+		{FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia14"), Device: true, Optional: true}}, {FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia15"), Device: true, Optional: true}},
+		{FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia16"), Device: true, Optional: true}}, {FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia17"), Device: true, Optional: true}},
+		{FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia18"), Device: true, Optional: true}}, {FilesystemConfig: &hst.FSBind{Source: fhs.AbsDev.Append("nvidia19"), Device: true, Optional: true}},
+	}...)
+}

cmd/hpkg/proc.go

@@ -0,0 +1,61 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"io"
+	"log"
+	"os"
+	"os/exec"
+
+	"hakurei.app/hst"
+	"hakurei.app/internal/info"
+	"hakurei.app/message"
+)
+
+var hakureiPathVal = info.MustHakureiPath().String()
+
+func mustRunApp(ctx context.Context, msg message.Msg, config *hst.Config, beforeFail func()) {
+	var (
+		cmd *exec.Cmd
+		st  io.WriteCloser
+	)
+
+	if r, w, err := os.Pipe(); err != nil {
+		beforeFail()
+		log.Fatalf("cannot pipe: %v", err)
+	} else {
+		if msg.IsVerbose() {
+			cmd = exec.CommandContext(ctx, hakureiPathVal, "-v", "app", "3")
+		} else {
+			cmd = exec.CommandContext(ctx, hakureiPathVal, "app", "3")
+		}
+		cmd.Stdin, cmd.Stdout, cmd.Stderr = os.Stdin, os.Stdout, os.Stderr
+		cmd.ExtraFiles = []*os.File{r}
+		st = w
+	}
+
+	go func() {
+		if err := json.NewEncoder(st).Encode(config); err != nil {
+			beforeFail()
+			log.Fatalf("cannot send configuration: %v", err)
+		}
+	}()
+
+	if err := cmd.Start(); err != nil {
+		beforeFail()
+		log.Fatalf("cannot start hakurei: %v", err)
+	}
+	if err := cmd.Wait(); err != nil {
+		var exitError *exec.ExitError
+		if errors.As(err, &exitError) {
+			beforeFail()
+			msg.BeforeExit()
+			os.Exit(exitError.ExitCode())
+		} else {
+			beforeFail()
+			log.Fatalf("cannot wait: %v", err)
+		}
+	}
+}

cmd/hpkg/test/configuration.nix

@@ -0,0 +1,62 @@
+{ 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 = {
+    variables = {
+      SWAYSOCK = "/tmp/sway-ipc.sock";
+      WLR_RENDERER = "pixman";
+    };
+  };
+
+  # Automatically configure and start Sway when logging in on tty1:
+  programs.bash.loginShellInit = ''
+    if [ "$(tty)" = "/dev/tty1" ]; then
+      set -e
+
+      mkdir -p ~/.config/sway
+      (sed s/Mod4/Mod1/ /etc/sway/config &&
+      echo 'output * bg ${pkgs.nixos-artwork.wallpapers.simple-light-gray.gnomeFilePath} fill' &&
+      echo 'output Virtual-1 res 1680x1050') > ~/.config/sway/config
+
+      sway --validate
+      systemd-cat --identifier=session sway && touch /tmp/sway-exit-ok
+    fi
+  '';
+
+  programs.sway.enable = true;
+
+  virtualisation = {
+    diskSize = 6 * 1024;
+
+    qemu.options = [
+      # Need to switch to a different GPU driver than the default one (-vga std) so that Sway can launch:
+      "-vga none -device virtio-gpu-pci"
+
+      # Increase zstd performance:
+      "-smp 8"
+    ];
+  };
+
+  environment.hakurei = {
+    enable = true;
+    stateDir = "/var/lib/hakurei";
+    users.alice = 0;
+
+    extraHomeConfig = {
+      home.stateVersion = "23.05";
+    };
+  };
+}

cmd/hpkg/test/default.nix

@@ -0,0 +1,34 @@
+{
+  testers,
+  callPackage,
+
+  system,
+  self,
+}:
+let
+  buildPackage = self.buildPackage.${system};
+in
+testers.nixosTest {
+  name = "hpkg";
+  nodes.machine = {
+    environment.etc = {
+      "foot.pkg".source = callPackage ./foot.nix { inherit buildPackage; };
+    };
+
+    imports = [
+      ./configuration.nix
+
+      self.nixosModules.hakurei
+      self.inputs.home-manager.nixosModules.home-manager
+    ];
+  };
+
+  # adapted from nixos sway integration tests
+
+  # testScriptWithTypes:49: error: Cannot call function of unknown type
+  #           (machine.succeed if succeed else machine.execute)(
+  #           ^
+  # Found 1 error in 1 file (checked 1 source file)
+  skipTypeCheck = true;
+  testScript = builtins.readFile ./test.py;
+}

cmd/hpkg/test/foot.nix

@@ -0,0 +1,48 @@
+{
+  lib,
+  buildPackage,
+  foot,
+  wayland-utils,
+  inconsolata,
+}:
+
+buildPackage {
+  name = "foot";
+  inherit (foot) version;
+
+  identity = 2;
+  id = "org.codeberg.dnkl.foot";
+
+  modules = [
+    {
+      home.packages = [
+        foot
+
+        # For wayland-info:
+        wayland-utils
+      ];
+    }
+  ];
+
+  nixosModules = [
+    {
+      # To help with OCR:
+      environment.etc."xdg/foot/foot.ini".text = lib.generators.toINI { } {
+        main = {
+          font = "inconsolata:size=14";
+        };
+        colors = rec {
+          foreground = "000000";
+          background = "ffffff";
+          regular2 = foreground;
+        };
+      };
+
+      fonts.packages = [ inconsolata ];
+    }
+  ];
+
+  script = ''
+    exec foot "$@"
+  '';
+}

cmd/hpkg/test/test.py

@@ -0,0 +1,110 @@
+import json
+import shlex
+
+q = shlex.quote
+NODE_GROUPS = ["nodes", "floating_nodes"]
+
+
+def swaymsg(command: str = "", succeed=True, type="command"):
+    assert command != "" or type != "command", "Must specify command or type"
+    shell = q(f"swaymsg -t {q(type)} -- {q(command)}")
+    with machine.nested(
+            f"sending swaymsg {shell!r}" + " (allowed to fail)" * (not succeed)
+    ):
+        ret = (machine.succeed if succeed else machine.execute)(
+            f"su - alice -c {shell}"
+        )
+
+    # execute also returns a status code, but disregard.
+    if not succeed:
+        _, ret = ret
+
+    if not succeed and not ret:
+        return None
+
+    parsed = json.loads(ret)
+    return parsed
+
+
+def walk(tree):
+    yield tree
+    for group in NODE_GROUPS:
+        for node in tree.get(group, []):
+            yield from walk(node)
+
+
+def wait_for_window(pattern):
+    def func(last_chance):
+        nodes = (node["name"] for node in walk(swaymsg(type="get_tree")))
+
+        if last_chance:
+            nodes = list(nodes)
+            machine.log(f"Last call! Current list of windows: {nodes}")
+
+        return any(pattern in name for name in nodes)
+
+    retry(func)
+
+
+def collect_state_ui(name):
+    swaymsg(f"exec hakurei ps > '/tmp/{name}.ps'")
+    machine.copy_from_vm(f"/tmp/{name}.ps", "")
+    swaymsg(f"exec hakurei --json ps > '/tmp/{name}.json'")
+    machine.copy_from_vm(f"/tmp/{name}.json", "")
+    machine.screenshot(name)
+
+
+def check_state(name, enablements):
+    instances = json.loads(machine.succeed("sudo -u alice -i XDG_RUNTIME_DIR=/run/user/1000 hakurei --json ps"))
+    if len(instances) != 1:
+        raise Exception(f"unexpected state length {len(instances)}")
+    instance = instances[0]
+
+    if len(instance['container']['args']) != 1 or not (instance['container']['args'][0].startswith("/nix/store/")) or f"hakurei-{name}-" not in (instance['container']['args'][0]):
+        raise Exception(f"unexpected args {instance['container']['args']}")
+
+    if instance['enablements'] != enablements:
+        raise Exception(f"unexpected enablements {instance['enablements']}")
+
+
+start_all()
+machine.wait_for_unit("multi-user.target")
+
+# To check hakurei's version:
+print(machine.succeed("sudo -u alice -i hakurei version"))
+
+# Wait for Sway to complete startup:
+machine.wait_for_file("/run/user/1000/wayland-1")
+machine.wait_for_file("/tmp/sway-ipc.sock")
+
+# Prepare hpkg directory:
+machine.succeed("install -dm 0700 -o alice -g users /var/lib/hakurei/1000")
+
+# Install hpkg app:
+swaymsg("exec hpkg -v install /etc/foot.pkg && touch /tmp/hpkg-install-ok")
+machine.wait_for_file("/tmp/hpkg-install-ok")
+
+# Start app (foot) with Wayland enablement:
+swaymsg("exec hpkg -v start org.codeberg.dnkl.foot")
+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, "pipewire": True})
+# Verify acl on XDG_RUNTIME_DIR:
+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 /tmp/hakurei.0/runtime | grep 10002")
+
+# Exit Sway and verify process exit status 0:
+swaymsg("exit", succeed=False)
+machine.wait_for_file("/tmp/sway-exit-ok")
+
+# Print hakurei share and rundir contents:
+print(machine.succeed("find /tmp/hakurei.0 "
+    + "-path '/tmp/hakurei.0/runtime/*/*' -prune -o "
+    + "-path '/tmp/hakurei.0/tmpdir/*/*' -prune -o "
+    + "-print"))
+print(machine.fail("ls /run/user/1000/hakurei"))

cmd/hpkg/with.go

@@ -0,0 +1,130 @@
+package main
+
+import (
+	"context"
+	"os"
+	"strings"
+
+	"hakurei.app/container/check"
+	"hakurei.app/container/fhs"
+	"hakurei.app/hst"
+	"hakurei.app/message"
+)
+
+func withNixDaemon(
+	ctx context.Context,
+	msg message.Msg,
+	action string, command []string, net bool, updateConfig func(config *hst.Config) *hst.Config,
+	app *appInfo, pathSet *appPathSet, dropShell bool, beforeFail func(),
+) {
+	flags := hst.FMultiarch | hst.FUserns // nix sandbox requires userns
+	if net {
+		flags |= hst.FHostNet
+	}
+	if dropShell {
+		flags |= hst.FTty
+	}
+
+	mustRunAppDropShell(ctx, msg, updateConfig(&hst.Config{
+		ID: app.ID,
+
+		ExtraPerms: []hst.ExtraPermConfig{
+			{Path: dataHome, Execute: true},
+			{Ensure: true, Path: pathSet.baseDir, Read: true, Write: true, Execute: true},
+		},
+
+		Identity: app.Identity,
+
+		Container: &hst.ContainerConfig{
+			Hostname: formatHostname(app.Name) + "-" + action,
+
+			Filesystem: []hst.FilesystemConfigJSON{
+				{FilesystemConfig: &hst.FSBind{Target: fhs.AbsEtc, Source: pathSet.cacheDir.Append("etc"), Special: true}},
+				{FilesystemConfig: &hst.FSBind{Source: pathSet.nixPath, Target: pathNix, Write: true}},
+				{FilesystemConfig: &hst.FSLink{Target: pathCurrentSystem, Linkname: app.CurrentSystem.String()}},
+				{FilesystemConfig: &hst.FSLink{Target: pathBin, Linkname: pathSwBin.String()}},
+				{FilesystemConfig: &hst.FSLink{Target: fhs.AbsUsrBin, Linkname: pathSwBin.String()}},
+				{FilesystemConfig: &hst.FSBind{Target: pathDataData.Append(app.ID), Source: pathSet.homeDir, Write: true, Ensure: true}},
+			},
+
+			Username: "hakurei",
+			Shell:    pathShell,
+			Home:     pathDataData.Append(app.ID),
+
+			Path: pathShell,
+			Args: []string{bash, "-lc", "rm -f /nix/var/nix/daemon-socket/socket && " +
+				// start nix-daemon
+				"nix-daemon --store / & " +
+				// wait for socket to appear
+				"(while [ ! -S /nix/var/nix/daemon-socket/socket ]; do sleep 0.01; done) && " +
+				// create directory so nix stops complaining
+				"mkdir -p /nix/var/nix/profiles/per-user/root/channels && " +
+				strings.Join(command, " && ") +
+				// terminate nix-daemon
+				" && pkill nix-daemon",
+			},
+
+			Flags: flags,
+		},
+	}), dropShell, beforeFail)
+}
+
+func withCacheDir(
+	ctx context.Context,
+	msg message.Msg,
+	action string, command []string, workDir *check.Absolute,
+	app *appInfo, pathSet *appPathSet, dropShell bool, beforeFail func(),
+) {
+	flags := hst.FMultiarch
+	if dropShell {
+		flags |= hst.FTty
+	}
+
+	mustRunAppDropShell(ctx, msg, &hst.Config{
+		ID: app.ID,
+
+		ExtraPerms: []hst.ExtraPermConfig{
+			{Path: dataHome, Execute: true},
+			{Ensure: true, Path: pathSet.baseDir, Read: true, Write: true, Execute: true},
+			{Path: workDir, Execute: true},
+		},
+
+		Identity: app.Identity,
+
+		Container: &hst.ContainerConfig{
+			Hostname: formatHostname(app.Name) + "-" + action,
+
+			Filesystem: []hst.FilesystemConfigJSON{
+				{FilesystemConfig: &hst.FSBind{Target: fhs.AbsEtc, Source: workDir.Append(fhs.Etc), Special: true}},
+				{FilesystemConfig: &hst.FSBind{Source: workDir.Append("nix"), Target: pathNix}},
+				{FilesystemConfig: &hst.FSLink{Target: pathCurrentSystem, Linkname: app.CurrentSystem.String()}},
+				{FilesystemConfig: &hst.FSLink{Target: pathBin, Linkname: pathSwBin.String()}},
+				{FilesystemConfig: &hst.FSLink{Target: fhs.AbsUsrBin, Linkname: pathSwBin.String()}},
+				{FilesystemConfig: &hst.FSBind{Source: workDir, Target: hst.AbsPrivateTmp.Append("bundle")}},
+				{FilesystemConfig: &hst.FSBind{Target: pathDataData.Append(app.ID, "cache"), Source: pathSet.cacheDir, Write: true, Ensure: true}},
+			},
+
+			Username: "nixos",
+			Shell:    pathShell,
+			Home:     pathDataData.Append(app.ID, "cache"),
+
+			Path: pathShell,
+			Args: []string{bash, "-lc", strings.Join(command, " && ")},
+
+			Flags: flags,
+		},
+	}, dropShell, beforeFail)
+}
+
+func mustRunAppDropShell(ctx context.Context, msg message.Msg, config *hst.Config, dropShell bool, beforeFail func()) {
+	if dropShell {
+		if config.Container != nil {
+			config.Container.Args = []string{bash, "-l"}
+		}
+		mustRunApp(ctx, msg, config, beforeFail)
+		beforeFail()
+		msg.BeforeExit()
+		os.Exit(0)
+	}
+	mustRunApp(ctx, msg, config, beforeFail)
+}

cmd/hsu/conf.go

@@ -1,7 +0,0 @@
-//go:build !rosa
-
-package main
-
-// hsuConfPath is an absolute pathname to the hsu configuration file. Its
-// contents are interpreted by parseConfig.
-const hsuConfPath = "/etc/hsurc"

cmd/hsu/config_rosa.go

@@ -1,7 +0,0 @@
-//go:build rosa
-
-package main
-
-// hsuConfPath is the pathname to the hsu configuration file, specific to
-// Rosa OS. Its contents are interpreted by parseConfig.
-const hsuConfPath = "/system/etc/hsurc"

cmd/hsu/hst.go

@@ -1,6 +1,6 @@
 package main
 
-/* keep in sync with hst */
+/* copied from hst and must never be changed */
 
 const (
 	userOffset = 100000

cmd/hsu/main.go

@@ -1,64 +1,13 @@
-// hsu starts the hakurei shim as the target subordinate user.
-//
-// The hsu program must be installed with the setuid and setgid bit set, and
-// owned by root. A configuration file must be installed at /etc/hsurc with
-// permission bits 0400, and owned by root. Each line of the file specifies a
-// hakurei userid to kernel uid mapping. A line consists of the decimal string
-// representation of the uid of the user wishing to start hakurei containers,
-// followed by a space, followed by the decimal string representation of its
-// userid. Duplicate uid entries are ignored, with the first occurrence taking
-// effect.
-//
-// For example, to map the kernel uid 1000 to the hakurei user id 0:
-//
-//	1000 0
-//
-// # Internals
-//
-// Hakurei and hsu holds pathnames pointing to each other set at link time. For
-// this reason, a distribution of hakurei has fixed installation prefix. Since
-// this program is never invoked by the user, behaviour described in the
-// following paragraphs are considered an internal detail and not covered by the
-// compatibility promise.
-//
-// After checking credentials, hsu checks via /proc/ the absolute pathname of
-// its parent process, and fails if it does not match the hakurei pathname set
-// at link time. This is not a security feature: the priv-side is considered
-// trusted, and this feature makes no attempt to address the racy nature of
-// querying /proc/, or debuggers attached to the parent process. Instead, this
-// aims to discourage misuse and reduce confusion if the user accidentally
-// stumbles upon this program. It also prevents accidental use of the incorrect
-// installation of hsu in some environments.
-//
-// Since target container environment variables are set up in shim via the
-// [container] infrastructure, the environment is used for parameters from the
-// parent process.
-//
-// HAKUREI_SHIM specifies a single byte between '3' and '9' representing the
-// setup pipe file descriptor. It is passed as is to the shim process and is the
-// only value in the environment of the shim process. Since hsurc is not
-// accessible to the parent process, leaving this unset causes hsu to print the
-// corresponding hakurei user id of the parent and terminate.
-//
-// HAKUREI_IDENTITY specifies the identity of the instance being started and is
-// used to produce the kernel uid alongside hakurei user id looked up from hsurc.
-//
-// HAKUREI_GROUPS specifies supplementary groups to inherit from the credentials
-// of the parent process in a ' ' separated list of decimal string
-// representations of gid. This has the unfortunate consequence of allowing
-// users mapped via hsurc to effectively drop group membership, so special care
-// must be taken to ensure this does not lead to an increase in access. This is
-// not applicable to Rosa OS since unsigned code execution is not permitted
-// outside hakurei containers, and is generally nonapplicable to the security
-// model of hakurei, where all untrusted code runs within containers.
 package main
 
+// minimise imports to avoid inadvertently calling init or global variable functions
+
 import (
 	"bytes"
 	"fmt"
 	"log"
 	"os"
-	"path/filepath"
+	"path"
 	"runtime"
 	"slices"
 	"strconv"
@@ -67,13 +16,10 @@ import (
 )
 
 const (
-	// envShim is the name of the environment variable holding a single byte
+	// envIdentity is the name of the environment variable holding a
-	// representing the shim setup pipe file descriptor.
+	// single byte representing the shim setup pipe file descriptor.
 	envShim = "HAKUREI_SHIM"
-	// envIdentity is the name of the environment variable holding a decimal
+	// envGroups holds a ' ' separated list of string representations of
-	// string representation of the current application identity.
-	envIdentity = "HAKUREI_IDENTITY"
-	// envGroups holds a ' ' separated list of decimal string representations of
 	// supplementary group gid. Membership requirements are enforced.
 	envGroups = "HAKUREI_GROUPS"
 )
@@ -89,6 +35,7 @@ func main() {
 
 	log.SetFlags(0)
 	log.SetPrefix("hsu: ")
+	log.SetOutput(os.Stderr)
 
 	if os.Geteuid() != 0 {
 		log.Fatal("this program must be owned by uid 0 and have the setuid bit set")
@@ -102,13 +49,13 @@ func main() {
 		log.Fatal("this program must not be started by root")
 	}
 
-	if !filepath.IsAbs(hakureiPath) {
+	if !path.IsAbs(hakureiPath) {
 		log.Fatal("this program is compiled incorrectly")
 		return
 	}
 
 	var toolPath string
-	pexe := filepath.Join("/proc", strconv.Itoa(os.Getppid()), "exe")
+	pexe := path.Join("/proc", strconv.Itoa(os.Getppid()), "exe")
 	if p, err := os.Readlink(pexe); err != nil {
 		log.Fatalf("cannot read parent executable path: %v", err)
 	} else if strings.HasSuffix(p, " (deleted)") {
@@ -152,6 +99,8 @@ func main() {
 		// last possible uid outcome
 		uidEnd = 999919999
 	)
+
+	// cast to int for use with library functions
 	uid := int(toUser(userid, identity))
 
 	// final bounds check to catch any bugs
@@ -187,6 +136,7 @@ func main() {
 	}
 
 	// careful! users in the allowlist is effectively allowed to drop groups via hsu
+
 	if err := syscall.Setresgid(uid, uid, uid); err != nil {
 		log.Fatalf("cannot set gid: %v", err)
 	}
@@ -196,21 +146,10 @@ func main() {
 	if err := syscall.Setresuid(uid, uid, uid); err != nil {
 		log.Fatalf("cannot set uid: %v", err)
 	}
-
+	if _, _, errno := syscall.AllThreadsSyscall(syscall.SYS_PRCTL, PR_SET_NO_NEW_PRIVS, 1, 0); errno != 0 {
-	if _, _, errno := syscall.AllThreadsSyscall(
-		syscall.SYS_PRCTL,
-		PR_SET_NO_NEW_PRIVS, 1,
-		0,
-	); errno != 0 {
 		log.Fatalf("cannot set no_new_privs flag: %s", errno.Error())
 	}
-
+	if err := syscall.Exec(toolPath, []string{"hakurei", "shim"}, []string{envShim + "=" + shimSetupFd}); err != nil {
-	if err := syscall.Exec(toolPath, []string{
-		"hakurei",
-		"shim",
-	}, []string{
-		envShim + "=" + shimSetupFd,
-	}); err != nil {
 		log.Fatalf("cannot start shim: %v", err)
 	}
 

cmd/hsu/parse.go

@@ -18,9 +18,8 @@ const (
 	useridEnd = useridStart + rangeSize - 1
 )
 
-// parseUint32Fast parses a string representation of an unsigned 32-bit integer
+// parseUint32Fast parses a string representation of an unsigned 32-bit integer value
-// value using the fast path only. This limits the range of values it is defined
+// using the fast path only. This limits the range of values it is defined in.
-// in but is perfectly adequate for this use case.
 func parseUint32Fast(s string) (uint32, error) {
 	sLen := len(s)
 	if sLen < 1 {
@@ -41,14 +40,12 @@ func parseUint32Fast(s string) (uint32, error) {
 	return n, nil
 }
 
-// parseConfig reads a list of allowed users from r until it encounters puid or
+// parseConfig reads a list of allowed users from r until it encounters puid or [io.EOF].
-// [io.EOF].
 //
-// Each line of the file specifies a hakurei userid to kernel uid mapping. A
+// Each line of the file specifies a hakurei userid to kernel uid mapping. A line consists
-// line consists of the string representation of the uid of the user wishing to
+// of the string representation of the uid of the user wishing to start hakurei containers,
-// start hakurei containers, followed by a space, followed by the string
+// followed by a space, followed by the string representation of its userid. Duplicate uid
-// representation of its userid. Duplicate uid entries are ignored, with the
+// entries are ignored, with the first occurrence taking effect.
-// first occurrence taking effect.
 //
 // All string representations are parsed by calling parseUint32Fast.
 func parseConfig(r io.Reader, puid uint32) (userid uint32, ok bool, err error) {
@@ -84,6 +81,10 @@ func parseConfig(r io.Reader, puid uint32) (userid uint32, ok bool, err error) {
 	return useridEnd + 1, false, s.Err()
 }
 
+// hsuConfPath is an absolute pathname to the hsu configuration file.
+// Its contents are interpreted by parseConfig.
+const hsuConfPath = "/etc/hsurc"
+
 // mustParseConfig calls parseConfig to interpret the contents of hsuConfPath,
 // terminating the program if an error is encountered, the syntax is incorrect,
 // or the current user is not authorised to use hsu because its uid is missing.
@@ -111,6 +112,10 @@ func mustParseConfig(puid int) (userid uint32) {
 	return
 }
 
+// envIdentity is the name of the environment variable holding a
+// string representation of the current application identity.
+var envIdentity = "HAKUREI_IDENTITY"
+
 // mustReadIdentity calls parseUint32Fast to interpret the value stored in envIdentity,
 // terminating the program if the value is not set, malformed, or out of bounds.
 func mustReadIdentity() uint32 {

cmd/irdump/main.go

@@ -0,0 +1,76 @@
+package main
+
+import (
+	"errors"
+	"log"
+	"os"
+
+	"hakurei.app/command"
+	"hakurei.app/internal/pkg"
+)
+
+func main() {
+	log.SetFlags(0)
+	log.SetPrefix("irdump: ")
+
+	var (
+		flagOutput string
+		flagReal   bool
+		flagHeader bool
+		flagForce  bool
+		flagRaw    bool
+	)
+	c := command.New(os.Stderr, log.Printf, "irdump", func(args []string) (err error) {
+		var input *os.File
+		if len(args) != 1 {
+			return errors.New("irdump requires 1 argument")
+		}
+		if input, err = os.Open(args[0]); err != nil {
+			return
+		}
+		defer input.Close()
+
+		var output *os.File
+		if flagOutput == "" {
+			output = os.Stdout
+		} else {
+			defer output.Close()
+			if output, err = os.Create(flagOutput); err != nil {
+				return
+			}
+		}
+
+		var out string
+		if out, err = pkg.Disassemble(input, flagReal, flagHeader, flagForce, flagRaw); err != nil {
+			return
+		}
+		if _, err = output.WriteString(out); err != nil {
+			return
+		}
+		return
+	}).Flag(
+		&flagOutput,
+		"o", command.StringFlag(""),
+		"Output file for asm (leave empty for stdout)",
+	).Flag(
+		&flagReal,
+		"r", command.BoolFlag(false),
+		"skip label generation; idents print real value",
+	).Flag(
+		&flagHeader,
+		"H", command.BoolFlag(false),
+		"display artifact headers",
+	).Flag(
+		&flagForce,
+		"f", command.BoolFlag(false),
+		"force display (skip validations)",
+	).Flag(
+		&flagRaw,
+		"R", command.BoolFlag(false),
+		"don't format output",
+	)
+
+	c.MustParse(os.Args[1:], func(err error) {
+		log.Fatal(err)
+	})
+}

cmd/mbf/main.go

@@ -1,42 +1,20 @@
-// The mbf program is a frontend for [hakurei.app/internal/rosa].
-//
-// This program is not covered by the compatibility promise. The command line
-// interface, available packages and their behaviour, and even the on-disk
-// format, may change at any time.
-//
-// # Name
-//
-// The name mbf stands for maiden's best friend, as a tribute to the DOOM source
-// port of [the same name]. This name is a placeholder and is subject to change.
-//
-// [the same name]: https://www.doomwiki.org/wiki/MBF
 package main
 
 import (
 	"context"
 	"errors"
 	"fmt"
-	"io"
 	"log"
 	"os"
 	"os/signal"
 	"path/filepath"
 	"runtime"
-	"strconv"
-	"strings"
-	"sync"
-	"sync/atomic"
 	"syscall"
-	"time"
 	"unique"
 
-	"hakurei.app/check"
 	"hakurei.app/command"
 	"hakurei.app/container"
-	"hakurei.app/container/seccomp"
+	"hakurei.app/container/check"
-	"hakurei.app/container/std"
-	"hakurei.app/ext"
-	"hakurei.app/fhs"
 	"hakurei.app/internal/pkg"
 	"hakurei.app/internal/rosa"
 	"hakurei.app/message"
@@ -69,37 +47,29 @@ func main() {
 	}()
 
 	var (
-		flagQuiet bool
+		flagQuiet  bool
-		flagCures int
+		flagCures  int
-		flagBase  string
+		flagBase   string
-		flagIdle  bool
+		flagTShift int
-
-		flagHostAbstract bool
 	)
 	c := command.New(os.Stderr, log.Printf, "mbf", func([]string) (err error) {
 		msg.SwapVerbose(!flagQuiet)
 
-		flagBase = os.ExpandEnv(flagBase)
-		if flagBase == "" {
-			flagBase = "cache"
-		}
-
 		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 {
-		var flags int
+			if flagTShift < 0 {
-		if flagIdle {
+				cache.SetThreshold(0)
-			flags |= pkg.CSchedIdle
+			} else if flagTShift > 31 {
+				cache.SetThreshold(1 << 31)
+			} else {
+				cache.SetThreshold(1 << flagTShift)
+			}
 		}
-		if flagHostAbstract {
-			flags |= pkg.CHostAbstract
-		}
-		cache, err = pkg.Open(ctx, msg, flags, flagCures, base)
-
 		return
 	}).Flag(
 		&flagQuiet,
@@ -111,19 +81,12 @@ func main() {
 		"Maximum number of dependencies to cure at any given time",
 	).Flag(
 		&flagBase,
-		"d", command.StringFlag("$MBF_CACHE_DIR"),
+		"d", command.StringFlag("cache"),
 		"Directory to store cured artifacts",
 	).Flag(
-		&flagIdle,
+		&flagTShift,
-		"sched-idle", command.BoolFlag(false),
+		"tshift", command.IntFlag(-1),
-		"Set SCHED_IDLE scheduling policy",
+		"Dependency graph size exponent, to the power of 2",
-	).Flag(
-		&flagHostAbstract,
-		"host-abstract", command.BoolFlag(
-			os.Getenv("MBF_HOST_ABSTRACT") != "",
-		),
-		"Do not restrict networked cure containers from connecting to host "+
-			"abstract UNIX sockets",
 	)
 
 	{
@@ -146,310 +109,50 @@ func main() {
 		)
 	}
 
-	{
-		var (
-			flagStatus bool
-			flagReport string
-		)
-		c.NewCommand(
-			"info",
-			"Display out-of-band metadata of an artifact",
-			func(args []string) (err error) {
-				if len(args) == 0 {
-					return errors.New("info requires at least 1 argument")
-				}
-
-				var r *rosa.Report
-				if flagReport != "" {
-					if r, err = rosa.OpenReport(flagReport); err != nil {
-						return err
-					}
-					defer func() {
-						if closeErr := r.Close(); err == nil {
-							err = closeErr
-						}
-					}()
-					defer r.HandleAccess(&err)()
-				}
-
-				for i, name := range args {
-					if p, ok := rosa.ResolveName(name); !ok {
-						return fmt.Errorf("unknown artifact %q", name)
-					} else {
-						var suffix string
-						if version := rosa.Std.Version(p); version != rosa.Unversioned {
-							suffix += "-" + version
-						}
-						fmt.Println("name        : " + name + suffix)
-
-						meta := rosa.GetMetadata(p)
-						fmt.Println("description : " + meta.Description)
-						if meta.Website != "" {
-							fmt.Println("website     : " +
-								strings.TrimSuffix(meta.Website, "/"))
-						}
-						if len(meta.Dependencies) > 0 {
-							fmt.Print("depends on  :")
-							for _, d := range meta.Dependencies {
-								s := rosa.GetMetadata(d).Name
-								if version := rosa.Std.Version(d); version != rosa.Unversioned {
-									s += "-" + version
-								}
-								fmt.Print(" " + s)
-							}
-							fmt.Println()
-						}
-
-						const statusPrefix = "status      : "
-						if flagStatus {
-							if r == nil {
-								var f io.ReadSeekCloser
-								f, err = cache.OpenStatus(rosa.Std.Load(p))
-								if err != nil {
-									if errors.Is(err, os.ErrNotExist) {
-										fmt.Println(
-											statusPrefix + "not yet cured",
-										)
-									} else {
-										return
-									}
-								} else {
-									fmt.Print(statusPrefix)
-									_, err = io.Copy(os.Stdout, f)
-									if err = errors.Join(err, f.Close()); err != nil {
-										return
-									}
-								}
-							} else {
-								status, n := r.ArtifactOf(cache.Ident(rosa.Std.Load(p)))
-								if status == nil {
-									fmt.Println(
-										statusPrefix + "not in report",
-									)
-								} else {
-									fmt.Println("size        :", n)
-									fmt.Print(statusPrefix)
-									if _, err = os.Stdout.Write(status); err != nil {
-										return
-									}
-								}
-							}
-						}
-
-						if i != len(args)-1 {
-							fmt.Println()
-						}
-					}
-				}
-				return nil
-			},
-		).
-			Flag(
-				&flagStatus,
-				"status", command.BoolFlag(false),
-				"Display cure status if available",
-			).
-			Flag(
-				&flagReport,
-				"report", command.StringFlag(""),
-				"Load cure status from this report file instead of cache",
-			)
-	}
-
 	c.NewCommand(
-		"report",
+		"stage3",
-		"Generate an artifact cure report for the current cache",
+		"Check for toolchain 3-stage non-determinism",
 		func(args []string) (err error) {
-			var w *os.File
+			_, _, _, stage1 := (rosa.Std - 2).NewLLVM()
-			switch len(args) {
+			_, _, _, stage2 := (rosa.Std - 1).NewLLVM()
-			case 0:
+			_, _, _, stage3 := rosa.Std.NewLLVM()
-				w = os.Stdout
+			var (
+				pathname *check.Absolute
+				checksum [2]unique.Handle[pkg.Checksum]
+			)
 
-			case 1:
+			if pathname, _, err = cache.Cure(stage1); err != nil {
-				if w, err = os.OpenFile(
+				return err
-					args[0],
+			}
-					os.O_CREATE|os.O_EXCL|syscall.O_WRONLY,
+			log.Println("stage1:", pathname)
-					0400,
+
-				); err != nil {
+			if pathname, checksum[0], err = cache.Cure(stage2); err != nil {
-					return
+				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(),
 				}
-				defer func() {
+			} else {
-					closeErr := w.Close()
+				log.Println(
-					if err == nil {
+					"stage2 is identical to stage3",
-						err = closeErr
+					"("+pkg.Encode(checksum[0].Value())+")",
-					}
+				)
-				}()
-
-			default:
-				return errors.New("report requires 1 argument")
 			}
-
+			return
-			if ext.Isatty(int(w.Fd())) {
-				return errors.New("output appears to be a terminal")
-			}
-			return rosa.WriteReport(msg, w, cache)
 		},
 	)
 
-	{
-		var flagJobs int
-		c.NewCommand("updates", command.UsageInternal, func([]string) error {
-			var (
-				errsMu sync.Mutex
-				errs   []error
-
-				n atomic.Uint64
-			)
-
-			w := make(chan rosa.PArtifact)
-			var wg sync.WaitGroup
-			for range max(flagJobs, 1) {
-				wg.Go(func() {
-					for p := range w {
-						meta := rosa.GetMetadata(p)
-						if meta.ID == 0 {
-							continue
-						}
-
-						v, err := meta.GetVersions(ctx)
-						if err != nil {
-							errsMu.Lock()
-							errs = append(errs, err)
-							errsMu.Unlock()
-							continue
-						}
-
-						if current, latest :=
-							rosa.Std.Version(p),
-							meta.GetLatest(v); current != latest {
-
-							n.Add(1)
-							log.Printf("%s %s < %s", meta.Name, current, latest)
-							continue
-						}
-
-						msg.Verbosef("%s is up to date", meta.Name)
-					}
-				})
-			}
-
-		done:
-			for i := range rosa.PresetEnd {
-				select {
-				case w <- rosa.PArtifact(i):
-					break
-				case <-ctx.Done():
-					break done
-				}
-			}
-			close(w)
-			wg.Wait()
-
-			if v := n.Load(); v > 0 {
-				errs = append(errs, errors.New(strconv.Itoa(int(v))+
-					" package(s) are out of date"))
-			}
-			return errors.Join(errs...)
-		}).
-			Flag(
-				&flagJobs,
-				"j", command.IntFlag(32),
-				"Maximum number of simultaneous connections",
-			)
-	}
-
 	{
 		var (
-			flagGentoo   string
+			flagDump string
-			flagChecksum string
-
-			flagStage0 bool
-		)
-		c.NewCommand(
-			"stage3",
-			"Check for toolchain 3-stage non-determinism",
-			func(args []string) (err error) {
-				t := rosa.Std
-				if flagGentoo != "" {
-					t -= 3 // magic number to discourage misuse
-
-					var checksum pkg.Checksum
-					if len(flagChecksum) != 0 {
-						if err = pkg.Decode(&checksum, flagChecksum); err != nil {
-							return
-						}
-					}
-					rosa.SetGentooStage3(flagGentoo, checksum)
-				}
-
-				_, _, _, stage1 := (t - 2).NewLLVM()
-				_, _, _, stage2 := (t - 1).NewLLVM()
-				_, _, _, stage3 := t.NewLLVM()
-				var (
-					pathname *check.Absolute
-					checksum [2]unique.Handle[pkg.Checksum]
-				)
-
-				if pathname, _, err = cache.Cure(stage1); err != nil {
-					return err
-				}
-				log.Println("stage1:", pathname)
-
-				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(),
-					}
-				} else {
-					log.Println(
-						"stage2 is identical to stage3",
-						"("+pkg.Encode(checksum[0].Value())+")",
-					)
-				}
-
-				if flagStage0 {
-					if pathname, _, err = cache.Cure(
-						t.Load(rosa.Stage0),
-					); err != nil {
-						return err
-					}
-					log.Println(pathname)
-				}
-
-				return
-			},
-		).
-			Flag(
-				&flagGentoo,
-				"gentoo", command.StringFlag(""),
-				"Bootstrap from a Gentoo stage3 tarball",
-			).
-			Flag(
-				&flagChecksum,
-				"checksum", command.StringFlag(""),
-				"Checksum of Gentoo stage3 tarball",
-			).
-			Flag(
-				&flagStage0,
-				"stage0", command.BoolFlag(false),
-				"Create bootstrap stage0 tarball",
-			)
-	}
-
-	{
-		var (
-			flagDump   string
-			flagEnter  bool
-			flagExport string
 		)
 		c.NewCommand(
 			"cure",
@@ -458,44 +161,15 @@ func main() {
 				if len(args) != 1 {
 					return errors.New("cure requires 1 argument")
 				}
-				p, ok := rosa.ResolveName(args[0])
+				if p, ok := rosa.ResolveName(args[0]); !ok {
-				if !ok {
+					return fmt.Errorf("unsupported artifact %q", args[0])
-					return fmt.Errorf("unknown artifact %q", args[0])
+				} else if flagDump == "" {
-				}
-
-				switch {
-				default:
 					pathname, _, err := cache.Cure(rosa.Std.Load(p))
-					if err != nil {
+					if err == nil {
-						return err
+						log.Println(pathname)
 					}
-					log.Println(pathname)
+					return err
-
+				} else {
-					if flagExport != "" {
-						msg.Verbosef("exporting %s to %s...", args[0], flagExport)
-
-						var f *os.File
-						if f, err = os.OpenFile(
-							flagExport,
-							os.O_WRONLY|os.O_CREATE|os.O_EXCL,
-							0400,
-						); err != nil {
-							return err
-						} else if _, err = pkg.Flatten(
-							os.DirFS(pathname.String()),
-							".",
-							f,
-						); err != nil {
-							_ = f.Close()
-							return err
-						} else if err = f.Close(); err != nil {
-							return err
-						}
-					}
-
-					return nil
-
-				case flagDump != "":
 					f, err := os.OpenFile(
 						flagDump,
 						os.O_WRONLY|os.O_CREATE|os.O_EXCL,
@@ -511,15 +185,6 @@ func main() {
 					}
 
 					return f.Close()
-
-				case flagEnter:
-					return cache.EnterExec(
-						ctx,
-						rosa.Std.Load(p),
-						true, os.Stdin, os.Stdout, os.Stderr,
-						rosa.AbsSystem.Append("bin", "mksh"),
-						"sh",
-					)
 				}
 			},
 		).
@@ -527,181 +192,13 @@ func main() {
 				&flagDump,
 				"dump", command.StringFlag(""),
 				"Write IR to specified pathname and terminate",
-			).
-			Flag(
-				&flagExport,
-				"export", command.StringFlag(""),
-				"Export cured artifact to specified pathname",
-			).
-			Flag(
-				&flagEnter,
-				"enter", command.BoolFlag(false),
-				"Enter cure container with an interactive shell",
 			)
 	}
 
-	{
-		var (
-			flagNet     bool
-			flagSession bool
-
-			flagWithToolchain bool
-		)
-		c.NewCommand(
-			"shell",
-			"Interactive shell in the specified Rosa OS environment",
-			func(args []string) error {
-				presets := make([]rosa.PArtifact, len(args))
-				for i, arg := range args {
-					p, ok := rosa.ResolveName(arg)
-					if !ok {
-						return fmt.Errorf("unknown artifact %q", arg)
-					}
-					presets[i] = p
-				}
-				root := make(pkg.Collect, 0, 6+len(args))
-				root = rosa.Std.AppendPresets(root, presets...)
-
-				if flagWithToolchain {
-					musl, compilerRT, runtimes, clang := (rosa.Std - 1).NewLLVM()
-					root = append(root, musl, compilerRT, runtimes, clang)
-				} else {
-					root = append(root, rosa.Std.Load(rosa.Musl))
-				}
-				root = append(root,
-					rosa.Std.Load(rosa.Mksh),
-					rosa.Std.Load(rosa.Toybox),
-				)
-
-				if _, _, err := cache.Cure(&root); err == nil {
-					return errors.New("unreachable")
-				} else if !pkg.IsCollected(err) {
-					return err
-				}
-
-				type cureRes struct {
-					pathname *check.Absolute
-					checksum unique.Handle[pkg.Checksum]
-				}
-				cured := make(map[pkg.Artifact]cureRes)
-				for _, a := range root {
-					pathname, checksum, err := cache.Cure(a)
-					if err != nil {
-						return err
-					}
-					cured[a] = cureRes{pathname, checksum}
-				}
-
-				layers := pkg.PromoteLayers(root, func(a pkg.Artifact) (
-					*check.Absolute,
-					unique.Handle[pkg.Checksum],
-				) {
-					res := cured[a]
-					return res.pathname, res.checksum
-				}, func(i int, d pkg.Artifact) {
-					r := pkg.Encode(cache.Ident(d).Value())
-					if s, ok := d.(fmt.Stringer); ok {
-						if name := s.String(); name != "" {
-							r += "-" + name
-						}
-					}
-					msg.Verbosef("promoted layer %d as %s", i, r)
-				})
-
-				z := container.New(ctx, msg)
-				z.WaitDelay = 3 * time.Second
-				z.SeccompPresets = pkg.SeccompPresets
-				z.SeccompFlags |= seccomp.AllowMultiarch
-				z.ParentPerm = 0700
-				z.HostNet = flagNet
-				z.RetainSession = flagSession
-				z.Hostname = "localhost"
-				z.Uid, z.Gid = (1<<10)-1, (1<<10)-1
-				z.Stdin, z.Stdout, z.Stderr = os.Stdin, os.Stdout, os.Stderr
-				if s, ok := os.LookupEnv("TERM"); ok {
-					z.Env = append(z.Env, "TERM="+s)
-				}
-
-				var tempdir *check.Absolute
-				if s, err := filepath.Abs(os.TempDir()); err != nil {
-					return err
-				} else if tempdir, err = check.NewAbs(s); err != nil {
-					return err
-				}
-
-				z.Dir = fhs.AbsRoot
-				z.Env = []string{
-					"SHELL=/system/bin/mksh",
-					"PATH=/system/bin",
-					"HOME=/",
-				}
-				z.Path = rosa.AbsSystem.Append("bin", "mksh")
-				z.Args = []string{"mksh"}
-				z.
-					OverlayEphemeral(fhs.AbsRoot, layers...).
-					Place(
-						fhs.AbsEtc.Append("hosts"),
-						[]byte("127.0.0.1 localhost\n"),
-					).
-					Place(
-						fhs.AbsEtc.Append("passwd"),
-						[]byte("media_rw:x:1023:1023::/:/system/bin/sh\n"+
-							"nobody:x:65534:65534::/proc/nonexistent:/system/bin/false\n"),
-					).
-					Place(
-						fhs.AbsEtc.Append("group"),
-						[]byte("media_rw:x:1023:\nnobody:x:65534:\n"),
-					).
-					Bind(tempdir, fhs.AbsTmp, std.BindWritable).
-					Proc(fhs.AbsProc).Dev(fhs.AbsDev, true)
-
-				if err := z.Start(); err != nil {
-					return err
-				}
-				if err := z.Serve(); err != nil {
-					return err
-				}
-				return z.Wait()
-			},
-		).
-			Flag(
-				&flagNet,
-				"net", command.BoolFlag(false),
-				"Share host net namespace",
-			).
-			Flag(
-				&flagSession,
-				"session", command.BoolFlag(true),
-				"Retain session",
-			).
-			Flag(
-				&flagWithToolchain,
-				"with-toolchain", command.BoolFlag(false),
-				"Include the stage2 LLVM toolchain",
-			)
-
-	}
-
-	c.Command(
-		"help",
-		"Show this help message",
-		func([]string) error { c.PrintHelp(); return nil },
-	)
-
 	c.MustParse(os.Args[1:], func(err error) {
 		if cache != nil {
 			cache.Close()
 		}
-		if w, ok := err.(interface{ Unwrap() []error }); !ok {
+		log.Fatal(err)
-			log.Fatal(err)
-		} else {
-			errs := w.Unwrap()
-			for i, e := range errs {
-				if i == len(errs)-1 {
-					log.Fatal(e)
-				}
-				log.Println(e)
-			}
-		}
 	})
 }

cmd/sharefs/fuse-operations.h

@@ -7,8 +7,8 @@
 #endif
 
 #define SHAREFS_MEDIA_RW_ID (1 << 10) - 1 /* owning gid presented to userspace */
-#define SHAREFS_PERM_DIR 0770             /* permission bits for directories presented to userspace */
+#define SHAREFS_PERM_DIR 0700             /* permission bits for directories presented to userspace */
-#define SHAREFS_PERM_REG 0660             /* permission bits for regular files 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 */

cmd/sharefs/fuse.go

@@ -19,21 +19,21 @@ import (
 	"encoding/gob"
 	"errors"
 	"fmt"
+	"io"
 	"log"
 	"os"
 	"os/exec"
 	"os/signal"
-	"path/filepath"
+	"path"
 	"runtime"
 	"runtime/cgo"
 	"strconv"
 	"syscall"
 	"unsafe"
 
-	"hakurei.app/check"
 	"hakurei.app/container"
+	"hakurei.app/container/check"
 	"hakurei.app/container/std"
-	"hakurei.app/fhs"
 	"hakurei.app/hst"
 	"hakurei.app/internal/helper/proc"
 	"hakurei.app/internal/info"
@@ -84,10 +84,7 @@ func destroySetup(private_data unsafe.Pointer) (ok bool) {
 }
 
 //export sharefs_init
-func sharefs_init(
+func sharefs_init(_ *C.struct_fuse_conn_info, cfg *C.struct_fuse_config) unsafe.Pointer {
-	_ *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)
@@ -105,11 +102,7 @@ func sharefs_init(
 	cfg.negative_timeout = 0
 
 	// all future filesystem operations happen through this dirfd
-	if fd, err := syscall.Open(
+	if fd, err := syscall.Open(setup.Source.String(), syscall.O_DIRECTORY|syscall.O_RDONLY|syscall.O_CLOEXEC, 0); err != nil {
-		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 {
@@ -144,9 +137,9 @@ func sharefs_destroy(private_data unsafe.Pointer) {
 func showHelp(args *fuseArgs) {
 	executableName := sharefsName
 	if args.argc > 0 {
-		executableName = filepath.Base(C.GoString(*args.argv))
+		executableName = path.Base(C.GoString(*args.argv))
 	} else if name, err := os.Executable(); err == nil {
-		executableName = filepath.Base(name)
+		executableName = path.Base(name)
 	}
 
 	fmt.Printf("usage: %s [options] <mountpoint>\n\n", executableName)
@@ -175,11 +168,8 @@ func parseOpts(args *fuseArgs, setup *setupState, log *log.Logger) (ok bool) {
 		// Decimal string representation of gid to set when running as root.
 		setgid *C.char
 
-		// Decimal string representation of open file descriptor to read
+		// Decimal string representation of open file descriptor to read setupState from.
-		// setupState from.
+		// This is an internal detail for containerisation and must not be specified directly.
-		//
-		// This is an internal detail for containerisation and must not be
-		// specified directly.
 		setup *C.char
 	}
 
@@ -262,8 +252,7 @@ func parseOpts(args *fuseArgs, setup *setupState, log *log.Logger) (ok bool) {
 	return true
 }
 
-// copyArgs returns a heap allocated copy of an argument slice in fuse_args
+// copyArgs returns a heap allocated copy of an argument slice in fuse_args representation.
-// representation.
 func copyArgs(s ...string) fuseArgs {
 	if len(s) == 0 {
 		return fuseArgs{argc: 0, argv: nil, allocated: 0}
@@ -279,7 +268,6 @@ func copyArgs(s ...string) fuseArgs {
 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))))
@@ -299,8 +287,8 @@ func _main(s ...string) (exitCode int) {
 	args := copyArgs(s...)
 	defer freeArgs(&args)
 
-	// this causes the kernel to enforce access control based on struct stat
+	// this causes the kernel to enforce access control based on
-	// populated by sharefs_getattr
+	// struct stat populated by sharefs_getattr
 	unsafeAddArgument(&args, "-odefault_permissions\x00")
 
 	var priv C.struct_sharefs_private
@@ -453,7 +441,12 @@ func _main(s ...string) (exitCode int) {
 		// keep fuse_parse_cmdline happy in the container
 		z.Tmpfs(check.MustAbs(container.Nonexistent), 1<<10, 0755)
 
-		z.Path = fhs.AbsProcSelfExe
+		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
@@ -464,19 +457,15 @@ func _main(s ...string) (exitCode int) {
 			z.Stdin, z.Stdout, z.Stderr = os.Stdin, os.Stdout, os.Stderr
 		}
 		z.Bind(z.Path, z.Path, 0)
-		setup.Fuse = int(proc.ExtraFileSlice(
+		setup.Fuse = int(proc.ExtraFileSlice(&z.ExtraFiles, os.NewFile(uintptr(C.fuse_session_fd(se)), "fuse")))
-			&z.ExtraFiles,
-			os.NewFile(uintptr(C.fuse_session_fd(se)), "fuse"),
-		))
 
-		var setupPipe [2]*os.File
+		var setupWriter io.WriteCloser
-		if r, w, err := os.Pipe(); err != nil {
+		if fd, w, err := container.Setup(&z.ExtraFiles); err != nil {
 			log.Println(err)
 			return 5
 		} else {
-			z.Args = append(z.Args, "-osetup="+strconv.Itoa(3+len(z.ExtraFiles)))
+			z.Args = append(z.Args, "-osetup="+strconv.Itoa(fd))
-			z.ExtraFiles = append(z.ExtraFiles, r)
+			setupWriter = w
-			setupPipe[0], setupPipe[1] = r, w
 		}
 
 		if err := z.Start(); err != nil {
@@ -487,9 +476,6 @@ func _main(s ...string) (exitCode int) {
 			}
 			return 5
 		}
-		if err := setupPipe[0].Close(); err != nil {
-			log.Println(err)
-		}
 		if err := z.Serve(); err != nil {
 			if m, ok := message.GetMessage(err); ok {
 				log.Println(m)
@@ -499,10 +485,10 @@ func _main(s ...string) (exitCode int) {
 			return 5
 		}
 
-		if err := gob.NewEncoder(setupPipe[1]).Encode(&setup); err != nil {
+		if err := gob.NewEncoder(setupWriter).Encode(&setup); err != nil {
 			log.Println(err)
 			return 5
-		} else if err = setupPipe[1].Close(); err != nil {
+		} else if err = setupWriter.Close(); err != nil {
 			log.Println(err)
 		}
 

cmd/sharefs/fuse_test.go

@@ -6,7 +6,7 @@ import (
 	"reflect"
 	"testing"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
 )
 
 func TestParseOpts(t *testing.T) {

cmd/sharefs/main.go

@@ -1,10 +1,3 @@
-// The sharefs FUSE filesystem is a permissionless shared filesystem.
-//
-// This filesystem is the primary means of file sharing between hakurei
-// application containers. It serves the same purpose in Rosa OS as /sdcard
-// does in AOSP.
-//
-// See help message for all available options.
 package main
 
 import (

cmd/sharefs/test/raceattr.go

@@ -1,122 +0,0 @@
-//go:build raceattr
-
-// The raceattr program reproduces vfs inode file attribute race.
-//
-// Even though libfuse high-level API presents the address of a struct stat
-// alongside struct fuse_context, file attributes are actually inherent to the
-// inode, instead of the specific call from userspace. The kernel implementation
-// in fs/fuse/xattr.c appears to make stale data in the inode (set by a previous
-// call) impossible or very unlikely to reach userspace via the stat family of
-// syscalls. However, when using default_permissions to have the VFS check
-// permissions, this race still happens, despite the resulting struct stat being
-// correct when overriding the check via capabilities otherwise.
-//
-// This program reproduces the failure, but because of its continuous nature, it
-// is provided independent of the vm integration test suite.
-package main
-
-import (
-	"context"
-	"flag"
-	"log"
-	"os"
-	"os/signal"
-	"runtime"
-	"sync"
-	"sync/atomic"
-	"syscall"
-)
-
-func newStatAs(
-	ctx context.Context, cancel context.CancelFunc,
-	n *atomic.Uint64, ok *atomic.Bool,
-	uid uint32, pathname string,
-	continuous bool,
-) func() {
-	return func() {
-		runtime.LockOSThread()
-		defer cancel()
-
-		if _, _, errno := syscall.Syscall(
-			syscall.SYS_SETUID, uintptr(uid),
-			0, 0,
-		); errno != 0 {
-			cancel()
-			log.Printf("cannot set uid to %d: %s", uid, errno)
-		}
-
-		var stat syscall.Stat_t
-		for {
-			if ctx.Err() != nil {
-				return
-			}
-
-			if err := syscall.Lstat(pathname, &stat); err != nil {
-				// SHAREFS_PERM_DIR not world executable, or
-				// SHAREFS_PERM_REG not world readable
-				if !continuous {
-					cancel()
-				}
-				ok.Store(true)
-				log.Printf("uid %d: %v", uid, err)
-			} else if stat.Uid != uid {
-				// appears to be unreachable
-				if !continuous {
-					cancel()
-				}
-				ok.Store(true)
-				log.Printf("got uid %d instead of %d", stat.Uid, uid)
-			}
-			n.Add(1)
-		}
-	}
-}
-
-func main() {
-	log.SetFlags(0)
-	log.SetPrefix("raceattr: ")
-
-	p := flag.String("target", "/sdcard/raceattr", "pathname of test file")
-	u0 := flag.Int("uid0", 1<<10-1, "first uid")
-	u1 := flag.Int("uid1", 1<<10-2, "second uid")
-	count := flag.Int("count", 1, "threads per uid")
-	continuous := flag.Bool("continuous", false, "keep running even after reproduce")
-	flag.Parse()
-
-	if os.Geteuid() != 0 {
-		log.Fatal("this program must run as root")
-	}
-
-	ctx, cancel := signal.NotifyContext(
-		context.Background(),
-		syscall.SIGINT,
-		syscall.SIGTERM,
-		syscall.SIGHUP,
-	)
-
-	if err := os.WriteFile(*p, nil, 0); err != nil {
-		log.Fatal(err)
-	}
-
-	var (
-		wg sync.WaitGroup
-
-		n  atomic.Uint64
-		ok atomic.Bool
-	)
-
-	if *count < 1 {
-		*count = 1
-	}
-	for range *count {
-		wg.Go(newStatAs(ctx, cancel, &n, &ok, uint32(*u0), *p, *continuous))
-		if *u1 >= 0 {
-			wg.Go(newStatAs(ctx, cancel, &n, &ok, uint32(*u1), *p, *continuous))
-		}
-	}
-
-	wg.Wait()
-	if !*continuous && ok.Load() {
-		log.Printf("reproduced after %d calls", n.Load())
-	}
-}

container/autoetc.go

@@ -4,13 +4,14 @@ import (
 	"encoding/gob"
 	"fmt"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
-	"hakurei.app/fhs"
+	"hakurei.app/container/fhs"
 )
 
 func init() { gob.Register(new(AutoEtcOp)) }
 
-// Etc is a helper for appending [AutoEtcOp] to [Ops].
+// Etc appends an [Op] that expands host /etc into a toplevel symlink mirror with /etc semantics.
+// This is not a generic setup op. It is implemented here to reduce ipc overhead.
 func (f *Ops) Etc(host *check.Absolute, prefix string) *Ops {
 	e := &AutoEtcOp{prefix}
 	f.Mkdir(fhs.AbsEtc, 0755)
@@ -19,9 +20,6 @@ func (f *Ops) Etc(host *check.Absolute, prefix string) *Ops {
 	return f
 }
 
-// AutoEtcOp expands host /etc into a toplevel symlink mirror with /etc semantics.
-//
-// This is not a generic setup op. It is implemented here to reduce ipc overhead.
 type AutoEtcOp struct{ Prefix string }
 
 func (e *AutoEtcOp) Valid() bool                                { return e != nil }

container/autoetc_test.go

@@ -5,8 +5,8 @@ import (
 	"os"
 	"testing"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
-	"hakurei.app/internal/stub"
+	"hakurei.app/container/stub"
 )
 
 func TestAutoEtcOp(t *testing.T) {

container/autoroot.go

@@ -4,22 +4,20 @@ import (
 	"encoding/gob"
 	"fmt"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
-	"hakurei.app/fhs"
+	"hakurei.app/container/fhs"
 	"hakurei.app/message"
 )
 
 func init() { gob.Register(new(AutoRootOp)) }
 
-// Root is a helper for appending [AutoRootOp] to [Ops].
+// Root appends an [Op] that expands a directory into a toplevel bind mount mirror on container root.
+// This is not a generic setup op. It is implemented here to reduce ipc overhead.
 func (f *Ops) Root(host *check.Absolute, flags int) *Ops {
 	*f = append(*f, &AutoRootOp{host, flags, nil})
 	return f
 }
 
-// AutoRootOp expands a directory into a toplevel bind mount mirror on container root.
-//
-// This is not a generic setup op. It is implemented here to reduce ipc overhead.
 type AutoRootOp struct {
 	Host *check.Absolute
 	// passed through to bindMount

container/autoroot_test.go

@@ -5,9 +5,9 @@ import (
 	"os"
 	"testing"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
 	"hakurei.app/container/std"
-	"hakurei.app/internal/stub"
+	"hakurei.app/container/stub"
 	"hakurei.app/message"
 )
 

container/capability.go

@@ -3,8 +3,6 @@ package container
 import (
 	"syscall"
 	"unsafe"
-
-	"hakurei.app/ext"
 )
 
 const (
@@ -52,16 +50,10 @@ func capset(hdrp *capHeader, datap *[2]capData) error {
 }
 
 // capBoundingSetDrop drops a capability from the calling thread's capability bounding set.
-func capBoundingSetDrop(cap uintptr) error {
+func capBoundingSetDrop(cap uintptr) error { return Prctl(syscall.PR_CAPBSET_DROP, cap, 0) }
-	return ext.Prctl(syscall.PR_CAPBSET_DROP, cap, 0)
-}
 
 // capAmbientClearAll clears the ambient capability set of the calling thread.
-func capAmbientClearAll() error {
+func capAmbientClearAll() error { return Prctl(PR_CAP_AMBIENT, PR_CAP_AMBIENT_CLEAR_ALL, 0) }
-	return ext.Prctl(PR_CAP_AMBIENT, PR_CAP_AMBIENT_CLEAR_ALL, 0)
-}
 
 // capAmbientRaise adds to the ambient capability set of the calling thread.
-func capAmbientRaise(cap uintptr) error {
+func capAmbientRaise(cap uintptr) error { return Prctl(PR_CAP_AMBIENT, PR_CAP_AMBIENT_RAISE, cap) }
-	return ext.Prctl(PR_CAP_AMBIENT, PR_CAP_AMBIENT_RAISE, cap)
-}

container/check/absolute.go

@@ -2,10 +2,10 @@
 package check
 
 import (
-	"encoding"
+	"encoding/json"
 	"errors"
 	"fmt"
-	"path/filepath"
+	"path"
 	"slices"
 	"strings"
 	"syscall"
@@ -30,16 +30,6 @@ func (e AbsoluteError) Is(target error) bool {
 // Absolute holds a pathname checked to be absolute.
 type Absolute struct{ pathname unique.Handle[string] }
 
-var (
-	_ encoding.TextAppender    = new(Absolute)
-	_ encoding.TextMarshaler   = new(Absolute)
-	_ encoding.TextUnmarshaler = new(Absolute)
-
-	_ encoding.BinaryAppender    = new(Absolute)
-	_ encoding.BinaryMarshaler   = new(Absolute)
-	_ encoding.BinaryUnmarshaler = new(Absolute)
-)
-
 // ok returns whether [Absolute] is not the zero value.
 func (a *Absolute) ok() bool { return a != nil && *a != (Absolute{}) }
 
@@ -71,7 +61,7 @@ func (a *Absolute) Is(v *Absolute) bool {
 
 // NewAbs checks pathname and returns a new [Absolute] if pathname is absolute.
 func NewAbs(pathname string) (*Absolute, error) {
-	if !filepath.IsAbs(pathname) {
+	if !path.IsAbs(pathname) {
 		return nil, AbsoluteError(pathname)
 	}
 	return unsafeAbs(pathname), nil
@@ -86,35 +76,46 @@ func MustAbs(pathname string) *Absolute {
 	}
 }
 
-// Append calls [filepath.Join] with [Absolute] as the first element.
+// Append calls [path.Join] with [Absolute] as the first element.
 func (a *Absolute) Append(elem ...string) *Absolute {
-	return unsafeAbs(filepath.Join(append([]string{a.String()}, elem...)...))
+	return unsafeAbs(path.Join(append([]string{a.String()}, elem...)...))
 }
 
-// Dir calls [filepath.Dir] with [Absolute] as its argument.
+// Dir calls [path.Dir] with [Absolute] as its argument.
-func (a *Absolute) Dir() *Absolute { return unsafeAbs(filepath.Dir(a.String())) }
+func (a *Absolute) Dir() *Absolute { return unsafeAbs(path.Dir(a.String())) }
 
-// AppendText appends the checked pathname.
+// GobEncode returns the checked pathname.
-func (a *Absolute) AppendText(data []byte) ([]byte, error) {
+func (a *Absolute) GobEncode() ([]byte, error) {
-	return append(data, a.String()...), nil
+	return []byte(a.String()), nil
 }
 
-// MarshalText returns the checked pathname.
+// GobDecode stores data if it represents an absolute pathname.
-func (a *Absolute) MarshalText() ([]byte, error) { return a.AppendText(nil) }
+func (a *Absolute) GobDecode(data []byte) error {
-
-// UnmarshalText stores data if it represents an absolute pathname.
-func (a *Absolute) UnmarshalText(data []byte) error {
 	pathname := string(data)
-	if !filepath.IsAbs(pathname) {
+	if !path.IsAbs(pathname) {
 		return AbsoluteError(pathname)
 	}
 	a.pathname = unique.Make(pathname)
 	return nil
 }
 
-func (a *Absolute) AppendBinary(data []byte) ([]byte, error) { return a.AppendText(data) }
+// MarshalJSON returns a JSON representation of the checked pathname.
-func (a *Absolute) MarshalBinary() ([]byte, error)           { return a.MarshalText() }
+func (a *Absolute) MarshalJSON() ([]byte, error) {
-func (a *Absolute) UnmarshalBinary(data []byte) error        { return a.UnmarshalText(data) }
+	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)
+	}
+	a.pathname = unique.Make(pathname)
+	return nil
+}
 
 // SortAbs calls [slices.SortFunc] for a slice of [Absolute].
 func SortAbs(x []*Absolute) {

container/check/absolute_test.go

@@ -11,12 +11,12 @@ import (
 	"testing"
 	_ "unsafe" // for go:linkname
 
-	. "hakurei.app/check"
+	. "hakurei.app/container/check"
 )
 
 // unsafeAbs returns check.Absolute on any string value.
 //
-//go:linkname unsafeAbs hakurei.app/check.unsafeAbs
+//go:linkname unsafeAbs hakurei.app/container/check.unsafeAbs
 func unsafeAbs(pathname string) *Absolute
 
 func TestAbsoluteError(t *testing.T) {
@@ -170,20 +170,20 @@ func TestCodecAbsolute(t *testing.T) {
 
 		{"good", MustAbs("/etc"),
 			nil,
-			"\t\x7f\x06\x01\x02\xff\x82\x00\x00\x00\b\xff\x80\x00\x04/etc",
+			"\t\x7f\x05\x01\x02\xff\x82\x00\x00\x00\b\xff\x80\x00\x04/etc",
-			",\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\x06\x01\x02\xff\x82\x00\x00\x00\x0f\xff\x84\x01\x04/etc\x01\xfc\xc0\xed\x00\x00\x00",
+			",\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\x04/etc\x01\xfc\xc0\xed\x00\x00\x00",
 
 			`"/etc"`, `{"val":"/etc","magic":3236757504}`},
 		{"not absolute", nil,
 			AbsoluteError("etc"),
-			"\t\x7f\x06\x01\x02\xff\x82\x00\x00\x00\a\xff\x80\x00\x03etc",
+			"\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\x06\x01\x02\xff\x82\x00\x00\x00\x0f\xff\x84\x01\x03etc\x01\xfb\x01\x81\xda\x00\x00\x00",
+			",\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",
 
 			`"etc"`, `{"val":"etc","magic":3236757504}`},
 		{"zero", nil,
 			new(AbsoluteError),
-			"\t\x7f\x06\x01\x02\xff\x82\x00\x00\x00\x04\xff\x80\x00\x00",
+			"\t\x7f\x05\x01\x02\xff\x82\x00\x00\x00\x04\xff\x80\x00\x00",
-			",\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\x06\x01\x02\xff\x82\x00\x00\x00\f\xff\x84\x01\x00\x01\xfb\x01\x81\xda\x00\x00\x00",
+			",\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\f\xff\x84\x01\x00\x01\xfb\x01\x81\xda\x00\x00\x00",
 			`""`, `{"val":"","magic":3236757504}`},
 	}
 
@@ -347,6 +347,15 @@ func TestCodecAbsolute(t *testing.T) {
 			})
 		})
 	}
+
+	t.Run("json passthrough", func(t *testing.T) {
+		t.Parallel()
+
+		wantErr := "invalid character ':' looking for beginning of value"
+		if err := new(Absolute).UnmarshalJSON([]byte(":3")); err == nil || err.Error() != wantErr {
+			t.Errorf("UnmarshalJSON: error = %v, want %s", err, wantErr)
+		}
+	})
 }
 
 func TestAbsoluteWrap(t *testing.T) {

container/check/overlay.go

@@ -11,8 +11,7 @@ const (
 	SpecialOverlayPath = ":"
 )
 
-// EscapeOverlayDataSegment escapes a string for formatting into the data
+// EscapeOverlayDataSegment escapes a string for formatting into the data argument of an overlay mount call.
-// argument of an overlay mount system call.
 func EscapeOverlayDataSegment(s string) string {
 	if s == "" {
 		return ""

container/check/overlay_test.go

@@ -3,7 +3,7 @@ package check_test
 import (
 	"testing"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
 )
 
 func TestEscapeOverlayDataSegment(t *testing.T) {

container/container.go

@@ -1,5 +1,4 @@
-// Package container implements unprivileged Linux containers with built-in
+// Package container implements unprivileged Linux containers with built-in support for syscall filtering.
-// support for syscall filtering.
 package container
 
 import (
@@ -16,12 +15,10 @@ import (
 	. "syscall"
 	"time"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
+	"hakurei.app/container/fhs"
 	"hakurei.app/container/seccomp"
 	"hakurei.app/container/std"
-	"hakurei.app/ext"
-	"hakurei.app/fhs"
-	"hakurei.app/internal/landlock"
 	"hakurei.app/message"
 )
 
@@ -29,6 +26,9 @@ const (
 	// CancelSignal is the signal expected by container init on context cancel.
 	// A custom [Container.Cancel] function must eventually deliver this signal.
 	CancelSignal = SIGUSR2
+
+	// Timeout for writing initParams to Container.setup.
+	initSetupTimeout = 5 * time.Second
 )
 
 type (
@@ -37,34 +37,24 @@ type (
 	Container struct {
 		// Whether the container init should stay alive after its parent terminates.
 		AllowOrphan bool
-		// Whether to set SchedPolicy and SchedPriority via sched_setscheduler(2).
-		SetScheduler bool
-		// Scheduling policy to set via sched_setscheduler(2).
-		SchedPolicy ext.SchedPolicy
-		// Scheduling priority to set via sched_setscheduler(2). The zero value
-		// implies the minimum value supported by the current SchedPolicy.
-		SchedPriority ext.Int
 		// Cgroup fd, nil to disable.
 		Cgroup *int
-		// ExtraFiles passed through to initial process in the container, with
+		// ExtraFiles passed through to initial process in the container,
-		// behaviour identical to its [exec.Cmd] counterpart.
+		// with behaviour identical to its [exec.Cmd] counterpart.
 		ExtraFiles []*os.File
 
-		// Write end of a pipe connected to the init to deliver [Params].
+		// param pipe for shim and init
-		setup [2]*os.File
+		setup *os.File
-		// Cancels the context passed to the underlying cmd.
+		// cancels cmd
 		cancel context.CancelFunc
-		// Closed after Wait returns. Keeps the spawning thread alive.
+		// closed after Wait returns
 		wait chan struct{}
 
 		Stdin  io.Reader
 		Stdout io.Writer
 		Stderr io.Writer
 
-		// Custom cancellation behaviour for the underlying [exec.Cmd]. Must
+		Cancel    func(cmd *exec.Cmd) error
-		// deliver [CancelSignal] before returning.
-		Cancel func(cmd *exec.Cmd) error
-		// Copied to the underlying [exec.Cmd].
 		WaitDelay time.Duration
 
 		cmd *exec.Cmd
@@ -184,24 +174,31 @@ var (
 	closeOnExecErr  error
 )
 
-// ensureCloseOnExec ensures all currently open file descriptors have the
+// ensureCloseOnExec ensures all currently open file descriptors have the syscall.FD_CLOEXEC flag set.
-// syscall.FD_CLOEXEC flag set.
+// This is only ran once as it is intended to handle files left open by the parent, and any file opened
-//
+// on this side should already have syscall.FD_CLOEXEC set.
-// This is only ran once as it is intended to handle files left open by the
-// parent, and any file opened on this side should already have
-// syscall.FD_CLOEXEC set.
 func ensureCloseOnExec() error {
-	closeOnExecOnce.Do(func() { closeOnExecErr = doCloseOnExec() })
+	closeOnExecOnce.Do(func() {
+		const fdPrefixPath = "/proc/self/fd/"
+
+		var entries []os.DirEntry
+		if entries, closeOnExecErr = os.ReadDir(fdPrefixPath); closeOnExecErr != nil {
+			return
+		}
+
+		var fd int
+		for _, ent := range entries {
+			if fd, closeOnExecErr = strconv.Atoi(ent.Name()); closeOnExecErr != nil {
+				break // not reached
+			}
+			CloseOnExec(fd)
+		}
+	})
 
 	if closeOnExecErr == nil {
 		return nil
 	}
-	return &StartError{
+	return &StartError{Fatal: true, Step: "set FD_CLOEXEC on all open files", Err: closeOnExecErr, Passthrough: true}
-		Fatal:       true,
-		Step:        "set FD_CLOEXEC on all open files",
-		Err:         closeOnExecErr,
-		Passthrough: true,
-	}
 }
 
 // Start starts the container init. The init process blocks until Serve is called.
@@ -285,16 +282,10 @@ func (p *Container) Start() error {
 	}
 
 	// place setup pipe before user supplied extra files, this is later restored by init
-	if r, w, err := os.Pipe(); err != nil {
+	if fd, f, err := Setup(&p.cmd.ExtraFiles); err != nil {
-		return &StartError{
+		return &StartError{true, "set up params stream", err, false, false}
-			Fatal: true,
-			Step:  "set up params stream",
-			Err:   err,
-		}
 	} else {
-		fd := 3 + len(p.cmd.ExtraFiles)
+		p.setup = f
-		p.cmd.ExtraFiles = append(p.cmd.ExtraFiles, r)
-		p.setup[0], p.setup[1] = r, w
 		p.cmd.Env = []string{setupEnv + "=" + strconv.Itoa(fd)}
 	}
 	p.cmd.ExtraFiles = append(p.cmd.ExtraFiles, p.ExtraFiles...)
@@ -304,63 +295,43 @@ func (p *Container) Start() error {
 		runtime.LockOSThread()
 		p.wait = make(chan struct{})
 
-		// setup depending on per-thread state must happen here
+		done <- func() error { // setup depending on per-thread state must happen here
-		done <- func() error {
+			// PR_SET_NO_NEW_PRIVS: depends on per-thread state but acts on all processes created from that thread
-			// PR_SET_NO_NEW_PRIVS: thread-directed but acts on all processes
+			if err := SetNoNewPrivs(); err != nil {
-			// created from the calling thread
+				return &StartError{true, "prctl(PR_SET_NO_NEW_PRIVS)", err, false, false}
-			if err := setNoNewPrivs(); err != nil {
-				return &StartError{
-					Fatal: true,
-					Step:  "prctl(PR_SET_NO_NEW_PRIVS)",
-					Err:   err,
-				}
 			}
 
 			// landlock: depends on per-thread state but acts on a process group
 			{
-				rulesetAttr := &landlock.RulesetAttr{
+				rulesetAttr := &RulesetAttr{Scoped: LANDLOCK_SCOPE_SIGNAL}
-					Scoped: landlock.LANDLOCK_SCOPE_SIGNAL,
-				}
 				if !p.HostAbstract {
-					rulesetAttr.Scoped |= landlock.LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET
+					rulesetAttr.Scoped |= LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET
 				}
 
-				if abi, err := landlock.GetABI(); err != nil {
+				if abi, err := LandlockGetABI(); err != nil {
-					if p.HostAbstract || !p.HostNet {
+					if p.HostAbstract {
-						// landlock can be skipped here as it restricts access
+						// landlock can be skipped here as it restricts access to resources
-						// to resources already covered by namespaces (pid, net)
+						// already covered by namespaces (pid)
 						goto landlockOut
 					}
-					return &StartError{Step: "get landlock ABI", Err: err}
+					return &StartError{false, "get landlock ABI", err, false, false}
 				} else if abi < 6 {
 					if p.HostAbstract {
 						// see above comment
 						goto landlockOut
 					}
-					return &StartError{
+					return &StartError{false, "kernel version too old for LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET", ENOSYS, true, false}
-						Step:   "kernel too old for LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET",
-						Err:    ENOSYS,
-						Origin: true,
-					}
 				} else {
 					p.msg.Verbosef("landlock abi version %d", abi)
 				}
 
 				if rulesetFd, err := rulesetAttr.Create(0); err != nil {
-					return &StartError{
+					return &StartError{true, "create landlock ruleset", err, false, false}
-						Fatal: true,
-						Step:  "create landlock ruleset",
-						Err:   err,
-					}
 				} else {
 					p.msg.Verbosef("enforcing landlock ruleset %s", rulesetAttr)
-					if err = landlock.RestrictSelf(rulesetFd, 0); err != nil {
+					if err = LandlockRestrictSelf(rulesetFd, 0); err != nil {
 						_ = Close(rulesetFd)
-						return &StartError{
+						return &StartError{true, "enforce landlock ruleset", err, false, false}
-							Fatal: true,
-							Step:  "enforce landlock ruleset",
-							Err:   err,
-						}
 					}
 					if err = Close(rulesetFd); err != nil {
 						p.msg.Verbosef("cannot close landlock ruleset: %v", err)
@@ -371,52 +342,9 @@ func (p *Container) Start() error {
 			landlockOut:
 			}
 
-			// sched_setscheduler: thread-directed but acts on all processes
-			// created from the calling thread
-			if p.SetScheduler {
-				if p.SchedPolicy < 0 || p.SchedPolicy > ext.SCHED_LAST {
-					return &StartError{
-						Fatal: false,
-						Step:  "set scheduling policy",
-						Err:   EINVAL,
-					}
-				}
-
-				var param schedParam
-				if priority, err := p.SchedPolicy.GetPriorityMin(); err != nil {
-					return &StartError{
-						Fatal: true,
-						Step:  "get minimum priority",
-						Err:   err,
-					}
-				} else {
-					param.priority = max(priority, p.SchedPriority)
-				}
-
-				p.msg.Verbosef(
-					"setting scheduling policy %s priority %d",
-					p.SchedPolicy, param.priority,
-				)
-				if err := schedSetscheduler(
-					0, // calling thread
-					p.SchedPolicy,
-					&param,
-				); err != nil {
-					return &StartError{
-						Fatal: true,
-						Step:  "set scheduling policy",
-						Err:   err,
-					}
-				}
-			}
-
 			p.msg.Verbose("starting container init")
 			if err := p.cmd.Start(); err != nil {
-				return &StartError{
+				return &StartError{false, "start container init", err, false, true}
-					Step:        "start container init",
-					Err:         err,
-					Passthrough: true,
-				}
 			}
 			return nil
 		}()
@@ -428,40 +356,21 @@ func (p *Container) Start() error {
 }
 
 // Serve serves [Container.Params] to the container init.
-//
 // Serve must only be called once.
-func (p *Container) Serve() (err error) {
+func (p *Container) Serve() error {
-	if p.setup[0] == nil || p.setup[1] == nil {
+	if p.setup == nil {
 		panic("invalid serve")
 	}
 
-	done := make(chan struct{})
+	setup := p.setup
-	defer func() {
+	p.setup = nil
-		if closeErr := p.setup[1].Close(); err == nil {
+	if err := setup.SetDeadline(time.Now().Add(initSetupTimeout)); err != nil {
-			err = closeErr
+		return &StartError{true, "set init pipe deadline", err, false, true}
-		}
-
-		if err != nil {
-			p.cancel()
-		}
-		close(done)
-		p.setup[0], p.setup[1] = nil, nil
-	}()
-	if err = p.setup[0].Close(); err != nil {
-		return &StartError{
-			Fatal:       true,
-			Step:        "close read end of init pipe",
-			Err:         err,
-			Passthrough: true,
-		}
 	}
 
 	if p.Path == nil {
-		return &StartError{
+		p.cancel()
-			Step:   "invalid executable pathname",
+		return &StartError{false, "invalid executable pathname", EINVAL, true, false}
-			Err:    EINVAL,
-			Origin: true,
-		}
 	}
 
 	// do not transmit nil
@@ -472,31 +381,21 @@ func (p *Container) Serve() (err error) {
 		p.SeccompRules = make([]std.NativeRule, 0)
 	}
 
-	t := time.Now().UTC()
+	err := gob.NewEncoder(setup).Encode(&initParams{
-	go func(f *os.File) {
-		select {
-		case <-p.ctx.Done():
-			if cancelErr := f.SetWriteDeadline(t); cancelErr != nil {
-				p.msg.Verbose(err)
-			}
-
-		case <-done:
-			p.msg.Verbose("setup payload took", time.Since(t))
-			return
-		}
-	}(p.setup[1])
-
-	return gob.NewEncoder(p.setup[1]).Encode(&initParams{
 		p.Params,
 		Getuid(),
 		Getgid(),
 		len(p.ExtraFiles),
 		p.msg.IsVerbose(),
 	})
+	_ = setup.Close()
+	if err != nil {
+		p.cancel()
+	}
+	return err
 }
 
-// Wait blocks until the container init process to exit and releases any
+// Wait waits for the container init process to exit and releases any resources associated with the [Container].
-// resources associated with the [Container].
 func (p *Container) Wait() error {
 	if p.cmd == nil || p.cmd.Process == nil {
 		return EINVAL
@@ -541,13 +440,11 @@ func (p *Container) StderrPipe() (r io.ReadCloser, err error) {
 }
 
 func (p *Container) String() string {
-	return fmt.Sprintf(
+	return fmt.Sprintf("argv: %q, filter: %v, rules: %d, flags: %#x, presets: %#x",
-		"argv: %q, filter: %v, rules: %d, flags: %#x, presets: %#x",
+		p.Args, !p.SeccompDisable, len(p.SeccompRules), int(p.SeccompFlags), int(p.SeccompPresets))
-		p.Args, !p.SeccompDisable, len(p.SeccompRules), int(p.SeccompFlags), int(p.SeccompPresets),
-	)
 }
 
-// ProcessState returns the address of os.ProcessState held by the underlying [exec.Cmd].
+// ProcessState returns the address to os.ProcessState held by the underlying [exec.Cmd].
 func (p *Container) ProcessState() *os.ProcessState {
 	if p.cmd == nil {
 		return nil
@@ -555,8 +452,7 @@ func (p *Container) ProcessState() *os.ProcessState {
 	return p.cmd.ProcessState
 }
 
-// New returns the address to a new instance of [Container]. This value requires
+// New returns the address to a new instance of [Container] that requires further initialisation before use.
-// further initialisation before use.
 func New(ctx context.Context, msg message.Msg) *Container {
 	if msg == nil {
 		msg = message.New(nil)
@@ -565,18 +461,12 @@ func New(ctx context.Context, msg message.Msg) *Container {
 	p := &Container{ctx: ctx, msg: msg, Params: Params{Ops: new(Ops)}}
 	c, cancel := context.WithCancel(ctx)
 	p.cancel = cancel
-	p.cmd = exec.CommandContext(c, fhs.ProcSelfExe)
+	p.cmd = exec.CommandContext(c, MustExecutable(msg))
 	return p
 }
 
 // NewCommand calls [New] and initialises the [Params.Path] and [Params.Args] fields.
-func NewCommand(
+func NewCommand(ctx context.Context, msg message.Msg, pathname *check.Absolute, name string, args ...string) *Container {
-	ctx context.Context,
-	msg message.Msg,
-	pathname *check.Absolute,
-	name string,
-	args ...string,
-) *Container {
 	z := New(ctx, msg)
 	z.Path = pathname
 	z.Args = append([]string{name}, args...)

container/container_test.go

@@ -16,21 +16,18 @@ import (
 	"strings"
 	"syscall"
 	"testing"
+	"time"
 
-	"hakurei.app/check"
 	"hakurei.app/command"
 	"hakurei.app/container"
+	"hakurei.app/container/check"
+	"hakurei.app/container/fhs"
 	"hakurei.app/container/seccomp"
 	"hakurei.app/container/std"
-	"hakurei.app/ext"
+	"hakurei.app/container/vfs"
-	"hakurei.app/fhs"
 	"hakurei.app/hst"
-	"hakurei.app/internal/info"
-	"hakurei.app/internal/landlock"
-	"hakurei.app/internal/params"
 	"hakurei.app/ldd"
 	"hakurei.app/message"
-	"hakurei.app/vfs"
 )
 
 // Note: this package requires cgo, which is unavailable in the Go playground.
@@ -86,9 +83,9 @@ func TestStartError(t *testing.T) {
 		{"params env", &container.StartError{
 			Fatal: true,
 			Step:  "set up params stream",
-			Err:   params.ErrReceiveEnv,
+			Err:   container.ErrReceiveEnv,
 		}, "set up params stream: environment variable not set",
-			params.ErrReceiveEnv, syscall.EBADF,
+			container.ErrReceiveEnv, syscall.EBADF,
 			"cannot set up params stream: environment variable not set"},
 
 		{"params", &container.StartError{
@@ -261,7 +258,7 @@ var containerTestCases = []struct {
 		1000, 100, nil, 0, std.PresetExt},
 	{"custom rules", true, true, true, false,
 		emptyOps, emptyMnt,
-		1, 31, []std.NativeRule{{Syscall: ext.SyscallNum(syscall.SYS_SETUID), Errno: std.ScmpErrno(syscall.EPERM)}}, 0, std.PresetExt},
+		1, 31, []std.NativeRule{{Syscall: std.ScmpSyscall(syscall.SYS_SETUID), Errno: std.ScmpErrno(syscall.EPERM)}}, 0, std.PresetExt},
 
 	{"tmpfs", true, false, false, true,
 		earlyOps(new(container.Ops).
@@ -438,8 +435,11 @@ func TestContainer(t *testing.T) {
 			wantOps, wantOpsCtx := tc.ops(t)
 			wantMnt := tc.mnt(t, wantOpsCtx)
 
+			ctx, cancel := context.WithTimeout(t.Context(), helperDefaultTimeout)
+			defer cancel()
+
 			var libPaths []*check.Absolute
-			c := helperNewContainerLibPaths(t.Context(), &libPaths, "container", strconv.Itoa(i))
+			c := helperNewContainerLibPaths(ctx, &libPaths, "container", strconv.Itoa(i))
 			c.Uid = tc.uid
 			c.Gid = tc.gid
 			c.Hostname = hostnameFromTestCase(tc.name)
@@ -449,6 +449,7 @@ func TestContainer(t *testing.T) {
 			} else {
 				c.Stdout, c.Stderr = os.Stdout, os.Stderr
 			}
+			c.WaitDelay = helperDefaultTimeout
 			*c.Ops = append(*c.Ops, *wantOps...)
 			c.SeccompRules = tc.rules
 			c.SeccompFlags = tc.flags | seccomp.AllowMultiarch
@@ -456,15 +457,6 @@ func TestContainer(t *testing.T) {
 			c.SeccompDisable = !tc.filter
 			c.RetainSession = tc.session
 			c.HostNet = tc.net
-			if info.CanDegrade {
-				if _, err := landlock.GetABI(); err != nil {
-					if !errors.Is(err, syscall.ENOSYS) {
-						t.Fatalf("LandlockGetABI: error = %v", err)
-					}
-					c.HostAbstract = true
-					t.Log("Landlock LSM is unavailable, enabling HostAbstract")
-				}
-			}
 
 			c.
 				Readonly(check.MustAbs(pathReadonly), 0755).
@@ -560,10 +552,11 @@ func testContainerCancel(
 ) func(t *testing.T) {
 	return func(t *testing.T) {
 		t.Parallel()
-		ctx, cancel := context.WithCancel(t.Context())
+		ctx, cancel := context.WithTimeout(t.Context(), helperDefaultTimeout)
 
 		c := helperNewContainer(ctx, "block")
 		c.Stdout, c.Stderr = os.Stdout, os.Stderr
+		c.WaitDelay = helperDefaultTimeout
 		if containerExtra != nil {
 			containerExtra(c)
 		}
@@ -744,7 +737,8 @@ func init() {
 const (
 	envDoCheck = "HAKUREI_TEST_DO_CHECK"
 
-	helperInnerPath = "/usr/bin/helper"
+	helperDefaultTimeout = 5 * time.Second
+	helperInnerPath      = "/usr/bin/helper"
 )
 
 var (
@@ -779,13 +773,14 @@ func TestMain(m *testing.M) {
 func helperNewContainerLibPaths(ctx context.Context, libPaths *[]*check.Absolute, args ...string) (c *container.Container) {
 	msg := message.New(nil)
 	msg.SwapVerbose(testing.Verbose())
+	executable := check.MustAbs(container.MustExecutable(msg))
 
 	c = container.NewCommand(ctx, msg, absHelperInnerPath, "helper", args...)
 	c.Env = append(c.Env, envDoCheck+"=1")
-	c.Bind(fhs.AbsProcSelfExe, absHelperInnerPath, 0)
+	c.Bind(executable, absHelperInnerPath, 0)
 
 	// in case test has cgo enabled
-	if entries, err := ldd.Resolve(ctx, msg, nil); err != nil {
+	if entries, err := ldd.Resolve(ctx, msg, executable); err != nil {
 		log.Fatalf("ldd: %v", err)
 	} else {
 		*libPaths = ldd.Path(entries)

container/dispatcher.go

@@ -1,10 +1,8 @@
 package container
 
 import (
-	"context"
 	"io"
 	"io/fs"
-	"net"
 	"os"
 	"os/exec"
 	"os/signal"
@@ -14,9 +12,6 @@ import (
 
 	"hakurei.app/container/seccomp"
 	"hakurei.app/container/std"
-	"hakurei.app/ext"
-	"hakurei.app/internal/netlink"
-	"hakurei.app/internal/params"
 	"hakurei.app/message"
 )
 
@@ -26,8 +21,7 @@ type osFile interface {
 	fs.File
 }
 
-// syscallDispatcher provides methods that make state-dependent system calls as
+// syscallDispatcher provides methods that make state-dependent system calls as part of their behaviour.
-// part of their behaviour.
 type syscallDispatcher interface {
 	// new starts a goroutine with a new instance of syscallDispatcher.
 	// A syscallDispatcher must never be used in any goroutine other than the one owning it,
@@ -57,7 +51,7 @@ type syscallDispatcher interface {
 	// isatty provides [Isatty].
 	isatty(fd int) bool
 	// receive provides [Receive].
-	receive(key string, e any, fdp *int) (closeFunc func() error, err error)
+	receive(key string, e any, fdp *uintptr) (closeFunc func() error, err error)
 
 	// bindMount provides procPaths.bindMount.
 	bindMount(msg message.Msg, source, target string, flags uintptr) error
@@ -68,7 +62,7 @@ type syscallDispatcher interface {
 	// ensureFile provides ensureFile.
 	ensureFile(name string, perm, pperm os.FileMode) error
 	// mustLoopback provides mustLoopback.
-	mustLoopback(ctx context.Context, msg message.Msg)
+	mustLoopback(msg message.Msg)
 
 	// seccompLoad provides [seccomp.Load].
 	seccompLoad(rules []std.NativeRule, flags seccomp.ExportFlag) error
@@ -146,18 +140,18 @@ func (k direct) new(f func(k syscallDispatcher)) { go f(k) }
 
 func (direct) lockOSThread() { runtime.LockOSThread() }
 
-func (direct) setPtracer(pid uintptr) error       { return ext.SetPtracer(pid) }
+func (direct) setPtracer(pid uintptr) error       { return SetPtracer(pid) }
-func (direct) setDumpable(dumpable uintptr) error { return ext.SetDumpable(dumpable) }
+func (direct) setDumpable(dumpable uintptr) error { return SetDumpable(dumpable) }
-func (direct) setNoNewPrivs() error               { return setNoNewPrivs() }
+func (direct) setNoNewPrivs() error               { return SetNoNewPrivs() }
 
 func (direct) lastcap(msg message.Msg) uintptr                 { return LastCap(msg) }
 func (direct) capset(hdrp *capHeader, datap *[2]capData) error { return capset(hdrp, datap) }
 func (direct) capBoundingSetDrop(cap uintptr) error            { return capBoundingSetDrop(cap) }
 func (direct) capAmbientClearAll() error                       { return capAmbientClearAll() }
 func (direct) capAmbientRaise(cap uintptr) error               { return capAmbientRaise(cap) }
-func (direct) isatty(fd int) bool                              { return ext.Isatty(fd) }
+func (direct) isatty(fd int) bool                              { return Isatty(fd) }
-func (direct) receive(key string, e any, fdp *int) (func() error, error) {
+func (direct) receive(key string, e any, fdp *uintptr) (func() error, error) {
-	return params.Receive(key, e, fdp)
+	return Receive(key, e, fdp)
 }
 
 func (direct) bindMount(msg message.Msg, source, target string, flags uintptr) error {
@@ -172,50 +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(ctx context.Context, msg message.Msg) {
+func (direct) mustLoopback(msg message.Msg) { mustLoopback(msg) }
-	var lo int
-	if ifi, err := net.InterfaceByName("lo"); err != nil {
-		msg.GetLogger().Fatalln(err)
-	} else {
-		lo = ifi.Index
-	}
-
-	c, err := netlink.DialRoute(0)
-	if err != nil {
-		msg.GetLogger().Fatalln(err)
-	}
-
-	must := func(err error) {
-		if err == nil {
-			return
-		}
-		if closeErr := c.Close(); closeErr != nil {
-			msg.Verbosef("cannot close RTNETLINK: %v", closeErr)
-		}
-
-		switch err.(type) {
-		case *os.SyscallError:
-			msg.GetLogger().Fatalf("cannot %v", err)
-
-		case syscall.Errno:
-			msg.GetLogger().Fatalf("RTNETLINK answers: %v", err)
-
-		default:
-			if err == context.DeadlineExceeded || err == context.Canceled {
-				msg.GetLogger().Fatalf("interrupted RTNETLINK operation")
-			}
-			msg.GetLogger().Fatal("RTNETLINK answers with malformed message")
-		}
-	}
-	must(c.SendNewaddrLo(ctx, uint32(lo)))
-	must(c.SendIfInfomsg(ctx, syscall.RTM_NEWLINK, 0, &syscall.IfInfomsg{
-		Family: syscall.AF_UNSPEC,
-		Index:  int32(lo),
-		Flags:  syscall.IFF_UP,
-		Change: syscall.IFF_UP,
-	}))
-	must(c.Close())
-}
 
 func (direct) seccompLoad(rules []std.NativeRule, flags seccomp.ExportFlag) error {
 	return seccomp.Load(rules, flags)

container/dispatcher_test.go

@@ -2,7 +2,6 @@ package container
 
 import (
 	"bytes"
-	"context"
 	"fmt"
 	"io"
 	"io/fs"
@@ -19,7 +18,7 @@ import (
 
 	"hakurei.app/container/seccomp"
 	"hakurei.app/container/std"
-	"hakurei.app/internal/stub"
+	"hakurei.app/container/stub"
 	"hakurei.app/message"
 )
 
@@ -239,11 +238,8 @@ func sliceAddr[S any](s []S) *[]S { return &s }
 
 func newCheckedFile(t *testing.T, name, wantData string, closeErr error) osFile {
 	f := &checkedOsFile{t: t, name: name, want: wantData, closeErr: closeErr}
-	// check happens in Close, and cleanup is not guaranteed to run, so relying
+	// check happens in Close, and cleanup is not guaranteed to run, so relying on it for sloppy implementations will cause sporadic test results
-	// on it for sloppy implementations will cause sporadic test results
+	f.cleanup = runtime.AddCleanup(f, func(name string) { f.t.Fatalf("checkedOsFile %s became unreachable without a call to Close", name) }, f.name)
-	f.cleanup = runtime.AddCleanup(f, func(name string) {
-		panic("checkedOsFile " + name + " became unreachable without a call to Close")
-	}, name)
 	return f
 }
 
@@ -390,7 +386,7 @@ func (k *kstub) isatty(fd int) bool {
 	return expect.Ret.(bool)
 }
 
-func (k *kstub) receive(key string, e any, fdp *int) (closeFunc func() error, err error) {
+func (k *kstub) receive(key string, e any, fdp *uintptr) (closeFunc func() error, err error) {
 	k.Helper()
 	expect := k.Expects("receive")
 
@@ -408,17 +404,10 @@ func (k *kstub) receive(key string, e any, fdp *int) (closeFunc func() error, er
 		}
 		return nil
 	}
-
-	// avoid changing test cases
-	var fdpComp *uintptr
-	if fdp != nil {
-		fdpComp = new(uintptr(*fdp))
-	}
-
 	err = expect.Error(
 		stub.CheckArg(k.Stub, "key", key, 0),
 		stub.CheckArgReflect(k.Stub, "e", e, 1),
-		stub.CheckArgReflect(k.Stub, "fdp", fdpComp, 2))
+		stub.CheckArgReflect(k.Stub, "fdp", fdp, 2))
 
 	// 3 is unused so stores params
 	if expect.Args[3] != nil {
@@ -433,7 +422,7 @@ func (k *kstub) receive(key string, e any, fdp *int) (closeFunc func() error, er
 	if expect.Args[4] != nil {
 		if v, ok := expect.Args[4].(uintptr); ok && v >= 3 {
 			if fdp != nil {
-				*fdp = int(v)
+				*fdp = v
 			}
 		}
 	}
@@ -476,7 +465,7 @@ func (k *kstub) ensureFile(name string, perm, pperm os.FileMode) error {
 		stub.CheckArg(k.Stub, "pperm", pperm, 2))
 }
 
-func (*kstub) mustLoopback(context.Context, message.Msg) { /* noop */ }
+func (*kstub) mustLoopback(message.Msg) { /* noop */ }
 
 func (k *kstub) seccompLoad(rules []std.NativeRule, flags seccomp.ExportFlag) error {
 	k.Helper()

container/errors.go

@@ -5,9 +5,9 @@ import (
 	"os"
 	"syscall"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
+	"hakurei.app/container/vfs"
 	"hakurei.app/message"
-	"hakurei.app/vfs"
 )
 
 // messageFromError returns a printable error message for a supported concrete type.
@@ -43,8 +43,7 @@ func messageFromError(err error) (m string, ok bool) {
 }
 
 // messagePrefix checks and prefixes the error message of a non-pointer error.
-// While this is usable for pointer errors, such use should be avoided as nil
+// While this is usable for pointer errors, such use should be avoided as nil check is omitted.
-// check is omitted.
 func messagePrefix[T error](prefix string, err error) (string, bool) {
 	var targetError T
 	if errors.As(err, &targetError) {

container/errors_test.go

@@ -8,9 +8,9 @@ import (
 	"syscall"
 	"testing"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
-	"hakurei.app/internal/stub"
+	"hakurei.app/container/stub"
-	"hakurei.app/vfs"
+	"hakurei.app/container/vfs"
 )
 
 func TestMessageFromError(t *testing.T) {

container/executable.go

@@ -0,0 +1,34 @@
+package container
+
+import (
+	"fmt"
+	"log"
+	"os"
+	"sync"
+
+	"hakurei.app/message"
+)
+
+var (
+	executable     string
+	executableOnce sync.Once
+)
+
+func copyExecutable(msg message.Msg) {
+	if name, err := os.Executable(); err != nil {
+		m := fmt.Sprintf("cannot read executable path: %v", err)
+		if msg != nil {
+			msg.BeforeExit()
+			msg.GetLogger().Fatal(m)
+		} else {
+			log.Fatal(m)
+		}
+	} else {
+		executable = name
+	}
+}
+
+func MustExecutable(msg message.Msg) string {
+	executableOnce.Do(func() { copyExecutable(msg) })
+	return executable
+}

container/executable_test.go

@@ -0,0 +1,18 @@
+package container_test
+
+import (
+	"os"
+	"testing"
+
+	"hakurei.app/container"
+	"hakurei.app/message"
+)
+
+func TestExecutable(t *testing.T) {
+	t.Parallel()
+	for i := 0; i < 16; i++ {
+		if got := container.MustExecutable(message.New(nil)); got != os.Args[0] {
+			t.Errorf("MustExecutable: %q, want %q", got, os.Args[0])
+		}
+	}
+}

container/fhs/abs.go

@@ -3,14 +3,14 @@ package fhs
 import (
 	_ "unsafe" // for go:linkname
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
 )
 
 /* constants in this file bypass abs check, be extremely careful when changing them! */
 
 // unsafeAbs returns check.Absolute on any string value.
 //
-//go:linkname unsafeAbs hakurei.app/check.unsafeAbs
+//go:linkname unsafeAbs hakurei.app/container/check.unsafeAbs
 func unsafeAbs(pathname string) *check.Absolute
 
 var (
@@ -42,8 +42,6 @@ var (
 	AbsDevShm = unsafeAbs(DevShm)
 	// AbsProc is [Proc] as [check.Absolute].
 	AbsProc = unsafeAbs(Proc)
-	// AbsProcSelfExe is [ProcSelfExe] as [check.Absolute].
-	AbsProcSelfExe = unsafeAbs(ProcSelfExe)
 	// AbsSys is [Sys] as [check.Absolute].
 	AbsSys = unsafeAbs(Sys)
 )

container/fhs/fhs.go

@@ -9,8 +9,7 @@ const (
 	// Tmp points to the place for small temporary files.
 	Tmp = "/tmp/"
 
-	// Run points to a "tmpfs" file system for system packages to place runtime
+	// Run points to a "tmpfs" file system for system packages to place runtime data, socket files, and similar.
-	// data, socket files, and similar.
 	Run = "/run/"
 	// RunUser points to a directory containing per-user runtime directories,
 	// each usually individually mounted "tmpfs" instances.
@@ -18,12 +17,10 @@ const (
 
 	// Usr points to vendor-supplied operating system resources.
 	Usr = "/usr/"
-	// UsrBin points to binaries and executables for user commands that shall
+	// UsrBin points to binaries and executables for user commands that shall appear in the $PATH search path.
-	// appear in the $PATH search path.
 	UsrBin = Usr + "bin/"
 
-	// Var points to persistent, variable system data. Writable during normal
+	// Var points to persistent, variable system data. Writable during normal system operation.
-	// system operation.
 	Var = "/var/"
 	// VarLib points to persistent system data.
 	VarLib = Var + "lib/"
@@ -32,20 +29,12 @@ const (
 
 	// Dev points to the root directory for device nodes.
 	Dev = "/dev/"
-	// DevShm is the place for POSIX shared memory segments, as created via
+	// DevShm is the place for POSIX shared memory segments, as created via shm_open(3).
-	// shm_open(3).
 	DevShm = "/dev/shm/"
-	// Proc points to a virtual kernel file system exposing the process list and
+	// Proc points to a virtual kernel file system exposing the process list and other functionality.
-	// other functionality.
 	Proc = "/proc/"
-	// ProcSys points to a hierarchy below /proc/ that exposes a number of
+	// ProcSys points to a hierarchy below /proc/ that exposes a number of kernel tunables.
-	// kernel tunables.
 	ProcSys = Proc + "sys/"
-	// ProcSelf resolves to the process's own /proc/pid directory.
+	// Sys points to a virtual kernel file system exposing discovered devices and other functionality.
-	ProcSelf = Proc + "self/"
-	// ProcSelfExe is a symbolic link to program pathname.
-	ProcSelfExe = ProcSelf + "exe"
-	// Sys points to a virtual kernel file system exposing discovered devices
-	// and other functionality.
 	Sys = "/sys/"
 )

container/init.go

@@ -7,8 +7,7 @@ import (
 	"log"
 	"os"
 	"os/exec"
-	"os/signal"
+	"path"
-	"path/filepath"
 	"slices"
 	"strconv"
 	"sync"
@@ -16,10 +15,8 @@ import (
 	. "syscall"
 	"time"
 
+	"hakurei.app/container/fhs"
 	"hakurei.app/container/seccomp"
-	"hakurei.app/ext"
-	"hakurei.app/fhs"
-	"hakurei.app/internal/params"
 	"hakurei.app/message"
 )
 
@@ -36,12 +33,12 @@ const (
 	- 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
+	It should be noted that none of this should become relevant at any point since the resulting
-	since the resulting intermediate root tmpfs should be effectively anonymous. */
+	intermediate root tmpfs should be effectively anonymous. */
 	intermediateHostPath = fhs.Proc + "self/fd"
 
-	// setupEnv is the name of the environment variable holding the string
+	// setupEnv is the name of the environment variable holding the string representation of
-	// representation of the read end file descriptor of the setup params pipe.
+	// the read end file descriptor of the setup params pipe.
 	setupEnv = "HAKUREI_SETUP"
 
 	// exitUnexpectedWait4 is the exit code if wait4 returns an unexpected errno.
@@ -62,8 +59,7 @@ type (
 		// 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
+		// prefix returns a log message prefix, and whether this Op prints no identifying message on its own.
-		// identifying message on its own.
 		prefix() (string, bool)
 
 		Is(op Op) bool
@@ -75,11 +71,9 @@ type (
 	setupState struct {
 		nonrepeatable uintptr
 
-		// Whether early reaping has concluded. Must only be accessed in the
+		// Whether early reaping has concluded. Must only be accessed in the wait4 loop.
-		// wait4 loop.
 		processConcluded bool
-		// Process to syscall.WaitStatus populated in the wait4 loop. Freed
+		// Process to syscall.WaitStatus populated in the wait4 loop. Freed after early reaping concludes.
-		// after early reaping concludes.
 		process map[int]WaitStatus
 		// Synchronises access to process.
 		processMu sync.RWMutex
@@ -148,46 +142,44 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 	}
 
 	var (
-		param      initParams
+		params      initParams
-		closeSetup func() error
+		closeSetup  func() error
-		setupFd    int
+		setupFd     uintptr
+		offsetSetup int
 	)
-	if f, err := k.receive(setupEnv, &param, &setupFd); err != nil {
+	if f, err := k.receive(setupEnv, &params, &setupFd); err != nil {
 		if errors.Is(err, EBADF) {
 			k.fatal(msg, "invalid setup descriptor")
 		}
-		if errors.Is(err, params.ErrReceiveEnv) {
+		if errors.Is(err, ErrReceiveEnv) {
 			k.fatal(msg, setupEnv+" not set")
 		}
 
 		k.fatalf(msg, "cannot decode init setup payload: %v", err)
 	} else {
-		if param.Ops == nil {
+		if params.Ops == nil {
 			k.fatal(msg, "invalid setup parameters")
 		}
-		if param.ParentPerm == 0 {
+		if params.ParentPerm == 0 {
-			param.ParentPerm = 0755
+			params.ParentPerm = 0755
 		}
 
-		msg.SwapVerbose(param.Verbose)
+		msg.SwapVerbose(params.Verbose)
 		msg.Verbose("received setup parameters")
 		closeSetup = f
+		offsetSetup = int(setupFd + 1)
 	}
 
-	if !param.HostNet {
+	if !params.HostNet {
-		ctx, cancel := signal.NotifyContext(context.Background(), CancelSignal,
+		k.mustLoopback(msg)
-			os.Interrupt, SIGTERM, SIGQUIT)
-		defer cancel() // for panics
-		k.mustLoopback(ctx, msg)
-		cancel()
 	}
 
 	// write uid/gid map here so parent does not need to set dumpable
-	if err := k.setDumpable(ext.SUID_DUMP_USER); err != nil {
+	if err := k.setDumpable(SUID_DUMP_USER); err != nil {
 		k.fatalf(msg, "cannot set SUID_DUMP_USER: %v", err)
 	}
 	if err := k.writeFile(fhs.Proc+"self/uid_map",
-		append([]byte{}, strconv.Itoa(param.Uid)+" "+strconv.Itoa(param.HostUid)+" 1\n"...),
+		append([]byte{}, strconv.Itoa(params.Uid)+" "+strconv.Itoa(params.HostUid)+" 1\n"...),
 		0); err != nil {
 		k.fatalf(msg, "%v", err)
 	}
@@ -197,17 +189,17 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 		k.fatalf(msg, "%v", err)
 	}
 	if err := k.writeFile(fhs.Proc+"self/gid_map",
-		append([]byte{}, strconv.Itoa(param.Gid)+" "+strconv.Itoa(param.HostGid)+" 1\n"...),
+		append([]byte{}, strconv.Itoa(params.Gid)+" "+strconv.Itoa(params.HostGid)+" 1\n"...),
 		0); err != nil {
 		k.fatalf(msg, "%v", err)
 	}
-	if err := k.setDumpable(ext.SUID_DUMP_DISABLE); err != nil {
+	if err := k.setDumpable(SUID_DUMP_DISABLE); err != nil {
 		k.fatalf(msg, "cannot set SUID_DUMP_DISABLE: %v", err)
 	}
 
 	oldmask := k.umask(0)
-	if param.Hostname != "" {
+	if params.Hostname != "" {
-		if err := k.sethostname([]byte(param.Hostname)); err != nil {
+		if err := k.sethostname([]byte(params.Hostname)); err != nil {
 			k.fatalf(msg, "cannot set hostname: %v", err)
 		}
 	}
@@ -220,15 +212,14 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 	}
 
 	ctx, cancel := context.WithCancel(context.Background())
-	state := &setupState{process: make(map[int]WaitStatus), Params: &param.Params, Msg: msg, Context: ctx}
+	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
+	this step is mostly for gathering information that would otherwise be difficult to obtain
-	difficult to obtain via library functions after pivot_root, and
+	via library functions after pivot_root, and implementations are expected to avoid changing
-	implementations are expected to avoid changing the state of the mount
+	the state of the mount namespace */
-	namespace */
+	for i, op := range *params.Ops {
-	for i, op := range *param.Ops {
 		if op == nil || !op.Valid() {
 			k.fatalf(msg, "invalid op at index %d", i)
 		}
@@ -267,11 +258,11 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 		k.fatalf(msg, "cannot enter intermediate root: %v", err)
 	}
 
-	/* apply is called right after pivot_root and entering the new root. This
+	/* apply is called right after pivot_root and entering the new root;
-	step sets up the container filesystem, and implementations are expected to
+	this step sets up the container filesystem, and implementations are expected to keep the host root
-	keep the host root and sysroot mount points intact but otherwise can do
+	and sysroot mount points intact but otherwise can do whatever they need to;
-	whatever they need to. Calling chdir is allowed but discouraged. */
+	chdir is allowed but discouraged */
-	for i, op := range *param.Ops {
+	for i, op := range *params.Ops {
 		// ops already checked during early setup
 		if prefix, ok := op.prefix(); ok {
 			msg.Verbosef("%s %s", prefix, op)
@@ -295,7 +286,7 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 
 	{
 		var fd int
-		if err := ext.IgnoringEINTR(func() (err error) {
+		if err := IgnoringEINTR(func() (err error) {
 			fd, err = k.open(fhs.Root, O_DIRECTORY|O_RDONLY, 0)
 			return
 		}); err != nil {
@@ -327,7 +318,7 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 		k.fatalf(msg, "cannot clear the ambient capability set: %v", err)
 	}
 	for i := uintptr(0); i <= lastcap; i++ {
-		if param.Privileged && i == CAP_SYS_ADMIN {
+		if params.Privileged && i == CAP_SYS_ADMIN {
 			continue
 		}
 		if err := k.capBoundingSetDrop(i); err != nil {
@@ -336,7 +327,7 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 	}
 
 	var keep [2]uint32
-	if param.Privileged {
+	if params.Privileged {
 		keep[capToIndex(CAP_SYS_ADMIN)] |= capToMask(CAP_SYS_ADMIN)
 
 		if err := k.capAmbientRaise(CAP_SYS_ADMIN); err != nil {
@@ -350,13 +341,13 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 		k.fatalf(msg, "cannot capset: %v", err)
 	}
 
-	if !param.SeccompDisable {
+	if !params.SeccompDisable {
-		rules := param.SeccompRules
+		rules := params.SeccompRules
 		if len(rules) == 0 { // non-empty rules slice always overrides presets
-			msg.Verbosef("resolving presets %#x", param.SeccompPresets)
+			msg.Verbosef("resolving presets %#x", params.SeccompPresets)
-			rules = seccomp.Preset(param.SeccompPresets, param.SeccompFlags)
+			rules = seccomp.Preset(params.SeccompPresets, params.SeccompFlags)
 		}
-		if err := k.seccompLoad(rules, param.SeccompFlags); err != nil {
+		if err := k.seccompLoad(rules, params.SeccompFlags); err != nil {
 			// this also indirectly asserts PR_SET_NO_NEW_PRIVS
 			k.fatalf(msg, "cannot load syscall filter: %v", err)
 		}
@@ -365,10 +356,10 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 		msg.Verbose("syscall filter not configured")
 	}
 
-	extraFiles := make([]*os.File, param.Count)
+	extraFiles := make([]*os.File, params.Count)
 	for i := range extraFiles {
 		// setup fd is placed before all extra files
-		extraFiles[i] = k.newFile(uintptr(setupFd+1+i), "extra file "+strconv.Itoa(i))
+		extraFiles[i] = k.newFile(uintptr(offsetSetup+i), "extra file "+strconv.Itoa(i))
 	}
 	k.umask(oldmask)
 
@@ -446,7 +437,7 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 
 	// called right before startup of initial process, all state changes to the
 	// current process is prohibited during late
-	for i, op := range *param.Ops {
+	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 {
@@ -467,14 +458,14 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 		k.fatalf(msg, "cannot close setup pipe: %v", err)
 	}
 
-	cmd := exec.Command(param.Path.String())
+	cmd := exec.Command(params.Path.String())
 	cmd.Stdin, cmd.Stdout, cmd.Stderr = os.Stdin, os.Stdout, os.Stderr
-	cmd.Args = param.Args
+	cmd.Args = params.Args
-	cmd.Env = param.Env
+	cmd.Env = params.Env
 	cmd.ExtraFiles = extraFiles
-	cmd.Dir = param.Dir.String()
+	cmd.Dir = params.Dir.String()
 
-	msg.Verbosef("starting initial process %s", param.Path)
+	msg.Verbosef("starting initial process %s", params.Path)
 	if err := k.start(cmd); err != nil {
 		k.fatalf(msg, "%v", err)
 	}
@@ -492,9 +483,9 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 	for {
 		select {
 		case s := <-sig:
-			if s == CancelSignal && param.ForwardCancel && cmd.Process != nil {
+			if s == CancelSignal && params.ForwardCancel && cmd.Process != nil {
 				msg.Verbose("forwarding context cancellation")
-				if err := k.signal(cmd, os.Interrupt); err != nil && !errors.Is(err, os.ErrProcessDone) {
+				if err := k.signal(cmd, os.Interrupt); err != nil {
 					k.printf(msg, "cannot forward cancellation: %v", err)
 				}
 				continue
@@ -524,7 +515,7 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 				cancel()
 
 				// start timeout early
-				go func() { time.Sleep(param.AdoptWaitDelay); close(timeout) }()
+				go func() { time.Sleep(params.AdoptWaitDelay); close(timeout) }()
 
 				// close initial process files; this also keeps them alive
 				for _, f := range extraFiles {
@@ -568,7 +559,7 @@ func TryArgv0(msg message.Msg) {
 		msg = message.New(log.Default())
 	}
 
-	if len(os.Args) > 0 && filepath.Base(os.Args[0]) == initName {
+	if len(os.Args) > 0 && path.Base(os.Args[0]) == initName {
 		Init(msg)
 		msg.BeforeExit()
 		os.Exit(0)

container/init_test.go

@@ -7,11 +7,10 @@ import (
 	"testing"
 	"time"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
 	"hakurei.app/container/seccomp"
 	"hakurei.app/container/std"
-	"hakurei.app/internal/params"
+	"hakurei.app/container/stub"
-	"hakurei.app/internal/stub"
 )
 
 func TestInitEntrypoint(t *testing.T) {
@@ -41,7 +40,7 @@ func TestInitEntrypoint(t *testing.T) {
 				call("lockOSThread", stub.ExpectArgs{}, nil, nil),
 				call("getpid", stub.ExpectArgs{}, 1, nil),
 				call("setPtracer", stub.ExpectArgs{uintptr(0)}, nil, nil),
-				call("receive", stub.ExpectArgs{"HAKUREI_SETUP", new(initParams), new(uintptr)}, nil, params.ErrReceiveEnv),
+				call("receive", stub.ExpectArgs{"HAKUREI_SETUP", new(initParams), new(uintptr)}, nil, ErrReceiveEnv),
 				call("fatal", stub.ExpectArgs{[]any{"HAKUREI_SETUP not set"}}, nil, nil),
 			},
 		}, nil},

container/initbind.go

@@ -6,22 +6,20 @@ import (
 	"os"
 	"syscall"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
 	"hakurei.app/container/std"
 )
 
 func init() { gob.Register(new(BindMountOp)) }
 
-// Bind is a helper for appending [BindMountOp] to [Ops].
+// Bind appends an [Op] that bind mounts host path [BindMountOp.Source] on container path [BindMountOp.Target].
 func (f *Ops) Bind(source, target *check.Absolute, flags int) *Ops {
 	*f = append(*f, &BindMountOp{nil, source, target, flags})
 	return f
 }
 
-// BindMountOp creates a bind mount from host path Source to container path Target.
+// BindMountOp bind mounts host path Source on container path Target.
-//
+// Note that Flags uses bits declared in this package and should not be set with constants in [syscall].
-// Note that Flags uses bits declared in the [std] package and should not be set
-// with constants in [syscall].
 type BindMountOp struct {
 	sourceFinal, Source, Target *check.Absolute
 

container/initbind_test.go

@@ -6,9 +6,9 @@ import (
 	"syscall"
 	"testing"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
 	"hakurei.app/container/std"
-	"hakurei.app/internal/stub"
+	"hakurei.app/container/stub"
 )
 
 func TestBindMountOp(t *testing.T) {

container/initdaemon.go

@@ -12,8 +12,8 @@ import (
 	"syscall"
 	"time"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
-	"hakurei.app/fhs"
+	"hakurei.app/container/fhs"
 )
 
 func init() { gob.Register(new(DaemonOp)) }
@@ -24,7 +24,8 @@ const (
 	daemonTimeout = 5 * time.Second
 )
 
-// Daemon is a helper for appending [DaemonOp] to [Ops].
+// 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

container/initdaemon_test.go

@@ -4,8 +4,8 @@ import (
 	"os"
 	"testing"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
-	"hakurei.app/internal/stub"
+	"hakurei.app/container/stub"
 	"hakurei.app/message"
 )
 

container/initdev.go

@@ -3,11 +3,11 @@ package container
 import (
 	"encoding/gob"
 	"fmt"
-	"path/filepath"
+	"path"
 	. "syscall"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
-	"hakurei.app/fhs"
+	"hakurei.app/container/fhs"
 )
 
 func init() { gob.Register(new(MountDevOp)) }
@@ -19,9 +19,7 @@ func (f *Ops) Dev(target *check.Absolute, mqueue bool) *Ops {
 }
 
 // DevWritable appends an [Op] that mounts a writable subset of host /dev.
-//
+// There is usually no good reason to write to /dev, so this should always be followed by a [RemountOp].
-// There is usually no good reason to write to /dev, so this should always be
-// followed by a [RemountOp].
 func (f *Ops) DevWritable(target *check.Absolute, mqueue bool) *Ops {
 	*f = append(*f, &MountDevOp{target, mqueue, true})
 	return f
@@ -46,7 +44,7 @@ func (d *MountDevOp) apply(state *setupState, k syscallDispatcher) error {
 	}
 
 	for _, name := range []string{"null", "zero", "full", "random", "urandom", "tty"} {
-		targetPath := filepath.Join(target, name)
+		targetPath := path.Join(target, name)
 		if err := k.ensureFile(targetPath, 0444, state.ParentPerm); err != nil {
 			return err
 		}
@@ -62,7 +60,7 @@ func (d *MountDevOp) apply(state *setupState, k syscallDispatcher) error {
 	for i, name := range []string{"stdin", "stdout", "stderr"} {
 		if err := k.symlink(
 			fhs.Proc+"self/fd/"+string(rune(i+'0')),
-			filepath.Join(target, name),
+			path.Join(target, name),
 		); err != nil {
 			return err
 		}
@@ -72,13 +70,13 @@ func (d *MountDevOp) apply(state *setupState, k syscallDispatcher) error {
 		{fhs.Proc + "kcore", "core"},
 		{"pts/ptmx", "ptmx"},
 	} {
-		if err := k.symlink(pair[0], filepath.Join(target, pair[1])); err != nil {
+		if err := k.symlink(pair[0], path.Join(target, pair[1])); err != nil {
 			return err
 		}
 	}
 
-	devShmPath := filepath.Join(target, "shm")
+	devShmPath := path.Join(target, "shm")
-	devPtsPath := filepath.Join(target, "pts")
+	devPtsPath := path.Join(target, "pts")
 	for _, name := range []string{devShmPath, devPtsPath} {
 		if err := k.mkdir(name, state.ParentPerm); err != nil {
 			return err
@@ -92,7 +90,7 @@ func (d *MountDevOp) apply(state *setupState, k syscallDispatcher) error {
 
 	if state.RetainSession {
 		if k.isatty(Stdout) {
-			consolePath := filepath.Join(target, "console")
+			consolePath := path.Join(target, "console")
 			if err := k.ensureFile(consolePath, 0444, state.ParentPerm); err != nil {
 				return err
 			}
@@ -110,7 +108,7 @@ func (d *MountDevOp) apply(state *setupState, k syscallDispatcher) error {
 	}
 
 	if d.Mqueue {
-		mqueueTarget := filepath.Join(target, "mqueue")
+		mqueueTarget := path.Join(target, "mqueue")
 		if err := k.mkdir(mqueueTarget, state.ParentPerm); err != nil {
 			return err
 		}

container/initdev_test.go

@@ -4,8 +4,8 @@ import (
 	"os"
 	"testing"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
-	"hakurei.app/internal/stub"
+	"hakurei.app/container/stub"
 )
 
 func TestMountDevOp(t *testing.T) {

container/initmkdir.go

@@ -5,12 +5,12 @@ import (
 	"fmt"
 	"os"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
 )
 
 func init() { gob.Register(new(MkdirOp)) }
 
-// Mkdir is a helper for appending [MkdirOp] to [Ops].
+// Mkdir appends an [Op] that creates a directory in the container filesystem.
 func (f *Ops) Mkdir(name *check.Absolute, perm os.FileMode) *Ops {
 	*f = append(*f, &MkdirOp{name, perm})
 	return f

container/initmkdir_test.go

@@ -4,8 +4,8 @@ import (
 	"os"
 	"testing"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
-	"hakurei.app/internal/stub"
+	"hakurei.app/container/stub"
 )
 
 func TestMkdirOp(t *testing.T) {

container/initoverlay.go

@@ -6,8 +6,8 @@ import (
 	"slices"
 	"strings"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
-	"hakurei.app/fhs"
+	"hakurei.app/container/fhs"
 )
 
 const (
@@ -54,11 +54,8 @@ func (e *OverlayArgumentError) Error() string {
 	}
 }
 
-// Overlay is a helper for appending [MountOverlayOp] to [Ops].
+// Overlay appends an [Op] that mounts the overlay pseudo filesystem on [MountOverlayOp.Target].
-func (f *Ops) Overlay(
+func (f *Ops) Overlay(target, state, work *check.Absolute, layers ...*check.Absolute) *Ops {
-	target, state, work *check.Absolute,
-	layers ...*check.Absolute,
-) *Ops {
 	*f = append(*f, &MountOverlayOp{
 		Target: target,
 		Lower:  layers,
@@ -68,12 +65,13 @@ func (f *Ops) Overlay(
 	return f
 }
 
-// OverlayEphemeral appends a [MountOverlayOp] with an ephemeral upperdir and workdir.
+// OverlayEphemeral appends an [Op] that mounts the overlay pseudo filesystem on [MountOverlayOp.Target]
+// with an ephemeral upperdir and workdir.
 func (f *Ops) OverlayEphemeral(target *check.Absolute, layers ...*check.Absolute) *Ops {
 	return f.Overlay(target, fhs.AbsRoot, nil, layers...)
 }
 
-// OverlayReadonly appends a readonly [MountOverlayOp].
+// OverlayReadonly appends an [Op] that mounts the overlay pseudo filesystem readonly on [MountOverlayOp.Target]
 func (f *Ops) OverlayReadonly(target *check.Absolute, layers ...*check.Absolute) *Ops {
 	return f.Overlay(target, nil, nil, layers...)
 }
@@ -84,34 +82,25 @@ type MountOverlayOp struct {
 
 	// Any filesystem, does not need to be on a writable filesystem.
 	Lower []*check.Absolute
-	// Formatted for [OptionOverlayLowerdir].
+	// formatted for [OptionOverlayLowerdir], resolved, prefixed and escaped during early
-	//
-	// Resolved, prefixed and escaped during early.
 	lower []string
-
 	// The upperdir is normally on a writable filesystem.
 	//
-	// If Work is nil and Upper holds the special value [fhs.AbsRoot], an
+	// If Work is nil and Upper holds the special value [fhs.AbsRoot],
-	// ephemeral upperdir and workdir will be set up.
+	// an ephemeral upperdir and workdir will be set up.
 	//
-	// If both Work and Upper are nil, upperdir and workdir is omitted and the
+	// If both Work and Upper are nil, upperdir and workdir is omitted and the overlay is mounted readonly.
-	// overlay is mounted readonly.
 	Upper *check.Absolute
-	// Formatted for [OptionOverlayUpperdir].
+	// formatted for [OptionOverlayUpperdir], resolved, prefixed and escaped during early
-	//
-	// Resolved, prefixed and escaped during early.
 	upper string
-
 	// The workdir needs to be an empty directory on the same filesystem as upperdir.
 	Work *check.Absolute
-	// Formatted for [OptionOverlayWorkdir].
+	// formatted for [OptionOverlayWorkdir], resolved, prefixed and escaped during early
-	//
-	// Resolved, prefixed and escaped during early.
 	work string
 
 	ephemeral bool
 
-	// Used internally for mounting to the intermediate root.
+	// used internally for mounting to the intermediate root
 	noPrefix bool
 }
 

container/initoverlay_test.go

@@ -5,8 +5,8 @@ import (
 	"os"
 	"testing"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
-	"hakurei.app/internal/stub"
+	"hakurei.app/container/stub"
 )
 
 func TestMountOverlayOp(t *testing.T) {

container/initplace.go

@@ -5,8 +5,8 @@ import (
 	"fmt"
 	"syscall"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
-	"hakurei.app/fhs"
+	"hakurei.app/container/fhs"
 )
 
 const (
@@ -16,7 +16,7 @@ const (
 
 func init() { gob.Register(new(TmpfileOp)) }
 
-// Place is a helper for appending [TmpfileOp] to [Ops].
+// Place appends an [Op] that places a file in container path [TmpfileOp.Path] containing [TmpfileOp.Data].
 func (f *Ops) Place(name *check.Absolute, data []byte) *Ops {
 	*f = append(*f, &TmpfileOp{name, data})
 	return f

container/initplace_test.go

@@ -4,8 +4,8 @@ import (
 	"os"
 	"testing"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
-	"hakurei.app/internal/stub"
+	"hakurei.app/container/stub"
 )
 
 func TestTmpfileOp(t *testing.T) {
@@ -21,7 +21,7 @@ func TestTmpfileOp(t *testing.T) {
 			Path: samplePath,
 			Data: sampleData,
 		}, nil, nil, []stub.Call{
-			call("createTemp", stub.ExpectArgs{"/", "tmp.*"}, (*checkedOsFile)(nil), stub.UniqueError(5)),
+			call("createTemp", stub.ExpectArgs{"/", "tmp.*"}, newCheckedFile(t, "tmp.32768", sampleDataString, nil), stub.UniqueError(5)),
 		}, stub.UniqueError(5)},
 
 		{"Write", &Params{ParentPerm: 0700}, &TmpfileOp{
@@ -35,14 +35,14 @@ func TestTmpfileOp(t *testing.T) {
 			Path: samplePath,
 			Data: sampleData,
 		}, nil, nil, []stub.Call{
-			call("createTemp", stub.ExpectArgs{"/", "tmp.*"}, newCheckedFile(t, "tmp.Close", sampleDataString, stub.UniqueError(3)), nil),
+			call("createTemp", stub.ExpectArgs{"/", "tmp.*"}, newCheckedFile(t, "tmp.32768", sampleDataString, stub.UniqueError(3)), nil),
 		}, stub.UniqueError(3)},
 
 		{"ensureFile", &Params{ParentPerm: 0700}, &TmpfileOp{
 			Path: samplePath,
 			Data: sampleData,
 		}, nil, nil, []stub.Call{
-			call("createTemp", stub.ExpectArgs{"/", "tmp.*"}, newCheckedFile(t, "tmp.ensureFile", sampleDataString, nil), nil),
+			call("createTemp", stub.ExpectArgs{"/", "tmp.*"}, newCheckedFile(t, "tmp.32768", sampleDataString, nil), nil),
 			call("ensureFile", stub.ExpectArgs{"/sysroot/etc/passwd", os.FileMode(0444), os.FileMode(0700)}, nil, stub.UniqueError(2)),
 		}, stub.UniqueError(2)},
 
@@ -50,29 +50,29 @@ func TestTmpfileOp(t *testing.T) {
 			Path: samplePath,
 			Data: sampleData,
 		}, nil, nil, []stub.Call{
-			call("createTemp", stub.ExpectArgs{"/", "tmp.*"}, newCheckedFile(t, "tmp.bindMount", sampleDataString, nil), nil),
+			call("createTemp", stub.ExpectArgs{"/", "tmp.*"}, newCheckedFile(t, "tmp.32768", sampleDataString, nil), nil),
 			call("ensureFile", stub.ExpectArgs{"/sysroot/etc/passwd", os.FileMode(0444), os.FileMode(0700)}, nil, nil),
-			call("bindMount", stub.ExpectArgs{"tmp.bindMount", "/sysroot/etc/passwd", uintptr(0x5), false}, nil, stub.UniqueError(1)),
+			call("bindMount", stub.ExpectArgs{"tmp.32768", "/sysroot/etc/passwd", uintptr(0x5), false}, nil, stub.UniqueError(1)),
 		}, stub.UniqueError(1)},
 
 		{"remove", &Params{ParentPerm: 0700}, &TmpfileOp{
 			Path: samplePath,
 			Data: sampleData,
 		}, nil, nil, []stub.Call{
-			call("createTemp", stub.ExpectArgs{"/", "tmp.*"}, newCheckedFile(t, "tmp.remove", sampleDataString, nil), nil),
+			call("createTemp", stub.ExpectArgs{"/", "tmp.*"}, newCheckedFile(t, "tmp.32768", sampleDataString, nil), nil),
 			call("ensureFile", stub.ExpectArgs{"/sysroot/etc/passwd", os.FileMode(0444), os.FileMode(0700)}, nil, nil),
-			call("bindMount", stub.ExpectArgs{"tmp.remove", "/sysroot/etc/passwd", uintptr(0x5), false}, nil, nil),
+			call("bindMount", stub.ExpectArgs{"tmp.32768", "/sysroot/etc/passwd", uintptr(0x5), false}, nil, nil),
-			call("remove", stub.ExpectArgs{"tmp.remove"}, nil, stub.UniqueError(0)),
+			call("remove", stub.ExpectArgs{"tmp.32768"}, nil, stub.UniqueError(0)),
 		}, stub.UniqueError(0)},
 
 		{"success", &Params{ParentPerm: 0700}, &TmpfileOp{
 			Path: samplePath,
 			Data: sampleData,
 		}, nil, nil, []stub.Call{
-			call("createTemp", stub.ExpectArgs{"/", "tmp.*"}, newCheckedFile(t, "tmp.success", sampleDataString, nil), nil),
+			call("createTemp", stub.ExpectArgs{"/", "tmp.*"}, newCheckedFile(t, "tmp.32768", sampleDataString, nil), nil),
 			call("ensureFile", stub.ExpectArgs{"/sysroot/etc/passwd", os.FileMode(0444), os.FileMode(0700)}, nil, nil),
-			call("bindMount", stub.ExpectArgs{"tmp.success", "/sysroot/etc/passwd", uintptr(0x5), false}, nil, nil),
+			call("bindMount", stub.ExpectArgs{"tmp.32768", "/sysroot/etc/passwd", uintptr(0x5), false}, nil, nil),
-			call("remove", stub.ExpectArgs{"tmp.success"}, nil, nil),
+			call("remove", stub.ExpectArgs{"tmp.32768"}, nil, nil),
 		}, nil},
 	})
 

container/initproc.go

@@ -5,12 +5,12 @@ import (
 	"fmt"
 	. "syscall"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
 )
 
 func init() { gob.Register(new(MountProcOp)) }
 
-// Proc is a helper for appending [MountProcOp] to [Ops].
+// Proc appends an [Op] that mounts a private instance of proc.
 func (f *Ops) Proc(target *check.Absolute) *Ops {
 	*f = append(*f, &MountProcOp{target})
 	return f

container/initproc_test.go

@@ -4,8 +4,8 @@ import (
 	"os"
 	"testing"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
-	"hakurei.app/internal/stub"
+	"hakurei.app/container/stub"
 )
 
 func TestMountProcOp(t *testing.T) {

container/initremount.go

@@ -4,12 +4,12 @@ import (
 	"encoding/gob"
 	"fmt"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
 )
 
 func init() { gob.Register(new(RemountOp)) }
 
-// Remount is a helper for appending [RemountOp] to [Ops].
+// Remount appends an [Op] that applies [RemountOp.Flags] on container path [RemountOp.Target].
 func (f *Ops) Remount(target *check.Absolute, flags uintptr) *Ops {
 	*f = append(*f, &RemountOp{target, flags})
 	return f

container/initremount_test.go

@@ -4,8 +4,8 @@ import (
 	"syscall"
 	"testing"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
-	"hakurei.app/internal/stub"
+	"hakurei.app/container/stub"
 )
 
 func TestRemountOp(t *testing.T) {

container/initsymlink.go

@@ -3,9 +3,9 @@ package container
 import (
 	"encoding/gob"
 	"fmt"
-	"path/filepath"
+	"path"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
 )
 
 func init() { gob.Register(new(SymlinkOp)) }
@@ -30,7 +30,7 @@ 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 !filepath.IsAbs(l.LinkName) {
+		if !path.IsAbs(l.LinkName) {
 			return check.AbsoluteError(l.LinkName)
 		}
 		if name, err := k.readlink(l.LinkName); err != nil {
@@ -44,7 +44,7 @@ func (l *SymlinkOp) early(_ *setupState, k syscallDispatcher) error {
 
 func (l *SymlinkOp) apply(state *setupState, k syscallDispatcher) error {
 	target := toSysroot(l.Target.String())
-	if err := k.mkdirAll(filepath.Dir(target), state.ParentPerm); err != nil {
+	if err := k.mkdirAll(path.Dir(target), state.ParentPerm); err != nil {
 		return err
 	}
 	return k.symlink(l.LinkName, target)

container/initsymlink_test.go

@@ -4,8 +4,8 @@ import (
 	"os"
 	"testing"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
-	"hakurei.app/internal/stub"
+	"hakurei.app/container/stub"
 )
 
 func TestSymlinkOp(t *testing.T) {

container/inittmpfs.go

@@ -8,7 +8,7 @@ import (
 	"strconv"
 	. "syscall"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
 )
 
 func init() { gob.Register(new(MountTmpfsOp)) }

container/inittmpfs_test.go

@@ -5,8 +5,8 @@ import (
 	"syscall"
 	"testing"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
-	"hakurei.app/internal/stub"
+	"hakurei.app/container/stub"
 )
 
 func TestMountTmpfsOp(t *testing.T) {

container/landlock.go

@@ -1,11 +1,11 @@
-package landlock
+package container
 
 import (
 	"strings"
 	"syscall"
 	"unsafe"
 
-	"hakurei.app/ext"
+	"hakurei.app/container/std"
 )
 
 // include/uapi/linux/landlock.h
@@ -14,11 +14,11 @@ const (
 	LANDLOCK_CREATE_RULESET_VERSION = 1 << iota
 )
 
-// AccessFS is bitmask of handled filesystem actions.
+// LandlockAccessFS is bitmask of handled filesystem actions.
-type AccessFS uint64
+type LandlockAccessFS uint64
 
 const (
-	LANDLOCK_ACCESS_FS_EXECUTE AccessFS = 1 << iota
+	LANDLOCK_ACCESS_FS_EXECUTE LandlockAccessFS = 1 << iota
 	LANDLOCK_ACCESS_FS_WRITE_FILE
 	LANDLOCK_ACCESS_FS_READ_FILE
 	LANDLOCK_ACCESS_FS_READ_DIR
@@ -38,8 +38,7 @@ const (
 	_LANDLOCK_ACCESS_FS_DELIM
 )
 
-// String returns a space-separated string of [AccessFS] flags.
+func (f LandlockAccessFS) String() string {
-func (f AccessFS) String() string {
 	switch f {
 	case LANDLOCK_ACCESS_FS_EXECUTE:
 		return "execute"
@@ -90,8 +89,8 @@ func (f AccessFS) String() string {
 		return "fs_ioctl_dev"
 
 	default:
-		var c []AccessFS
+		var c []LandlockAccessFS
-		for i := AccessFS(1); i < _LANDLOCK_ACCESS_FS_DELIM; i <<= 1 {
+		for i := LandlockAccessFS(1); i < _LANDLOCK_ACCESS_FS_DELIM; i <<= 1 {
 			if f&i != 0 {
 				c = append(c, i)
 			}
@@ -107,18 +106,17 @@ func (f AccessFS) String() string {
 	}
 }
 
-// AccessNet is bitmask of handled network actions.
+// LandlockAccessNet is bitmask of handled network actions.
-type AccessNet uint64
+type LandlockAccessNet uint64
 
 const (
-	LANDLOCK_ACCESS_NET_BIND_TCP AccessNet = 1 << iota
+	LANDLOCK_ACCESS_NET_BIND_TCP LandlockAccessNet = 1 << iota
 	LANDLOCK_ACCESS_NET_CONNECT_TCP
 
 	_LANDLOCK_ACCESS_NET_DELIM
 )
 
-// String returns a space-separated string of [AccessNet] flags.
+func (f LandlockAccessNet) String() string {
-func (f AccessNet) String() string {
 	switch f {
 	case LANDLOCK_ACCESS_NET_BIND_TCP:
 		return "bind_tcp"
@@ -127,8 +125,8 @@ func (f AccessNet) String() string {
 		return "connect_tcp"
 
 	default:
-		var c []AccessNet
+		var c []LandlockAccessNet
-		for i := AccessNet(1); i < _LANDLOCK_ACCESS_NET_DELIM; i <<= 1 {
+		for i := LandlockAccessNet(1); i < _LANDLOCK_ACCESS_NET_DELIM; i <<= 1 {
 			if f&i != 0 {
 				c = append(c, i)
 			}
@@ -144,18 +142,17 @@ func (f AccessNet) String() string {
 	}
 }
 
-// Scope is bitmask of scopes restricting a Landlock domain from accessing outside resources.
+// LandlockScope is bitmask of scopes restricting a Landlock domain from accessing outside resources.
-type Scope uint64
+type LandlockScope uint64
 
 const (
-	LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET Scope = 1 << iota
+	LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET LandlockScope = 1 << iota
 	LANDLOCK_SCOPE_SIGNAL
 
 	_LANDLOCK_SCOPE_DELIM
 )
 
-// String returns a space-separated string of [Scope] flags.
+func (f LandlockScope) String() string {
-func (f Scope) String() string {
 	switch f {
 	case LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET:
 		return "abstract_unix_socket"
@@ -164,8 +161,8 @@ func (f Scope) String() string {
 		return "signal"
 
 	default:
-		var c []Scope
+		var c []LandlockScope
-		for i := Scope(1); i < _LANDLOCK_SCOPE_DELIM; i <<= 1 {
+		for i := LandlockScope(1); i < _LANDLOCK_SCOPE_DELIM; i <<= 1 {
 			if f&i != 0 {
 				c = append(c, i)
 			}
@@ -184,15 +181,13 @@ func (f Scope) String() string {
 // RulesetAttr is equivalent to struct landlock_ruleset_attr.
 type RulesetAttr struct {
 	// Bitmask of handled filesystem actions.
-	HandledAccessFS AccessFS
+	HandledAccessFS LandlockAccessFS
 	// Bitmask of handled network actions.
-	HandledAccessNet AccessNet
+	HandledAccessNet LandlockAccessNet
-	// Bitmask of scopes restricting a Landlock domain from accessing outside
+	// Bitmask of scopes restricting a Landlock domain from accessing outside resources (e.g. IPCs).
-	// resources (e.g. IPCs).
+	Scoped LandlockScope
-	Scoped Scope
 }
 
-// String returns a user-facing description of [RulesetAttr].
 func (rulesetAttr *RulesetAttr) String() string {
 	if rulesetAttr == nil {
 		return "NULL"
@@ -213,7 +208,6 @@ func (rulesetAttr *RulesetAttr) String() string {
 	return strings.Join(elems, ", ")
 }
 
-// Create loads the ruleset into the kernel.
 func (rulesetAttr *RulesetAttr) Create(flags uintptr) (fd int, err error) {
 	var pointer, size uintptr
 	// NULL needed for abi version
@@ -222,13 +216,10 @@ func (rulesetAttr *RulesetAttr) Create(flags uintptr) (fd int, err error) {
 		size = unsafe.Sizeof(*rulesetAttr)
 	}
 
-	rulesetFd, _, errno := syscall.Syscall(
+	rulesetFd, _, errno := syscall.Syscall(std.SYS_LANDLOCK_CREATE_RULESET, pointer, size, flags)
-		ext.SYS_LANDLOCK_CREATE_RULESET,
-		pointer, size,
-		flags,
-	)
 	fd = int(rulesetFd)
 	err = errno
+
 	if fd < 0 {
 		return
 	}
@@ -239,19 +230,12 @@ func (rulesetAttr *RulesetAttr) Create(flags uintptr) (fd int, err error) {
 	return fd, nil
 }
 
-// GetABI returns the ABI version supported by the kernel.
+func LandlockGetABI() (int, error) {
-func GetABI() (int, error) {
 	return (*RulesetAttr)(nil).Create(LANDLOCK_CREATE_RULESET_VERSION)
 }
 
-// RestrictSelf applies a loaded ruleset to the calling thread.
+func LandlockRestrictSelf(rulesetFd int, flags uintptr) error {
-func RestrictSelf(rulesetFd int, flags uintptr) error {
+	r, _, errno := syscall.Syscall(std.SYS_LANDLOCK_RESTRICT_SELF, uintptr(rulesetFd), flags, 0)
-	r, _, errno := syscall.Syscall(
-		ext.SYS_LANDLOCK_RESTRICT_SELF,
-		uintptr(rulesetFd),
-		flags,
-		0,
-	)
 	if r != 0 {
 		return errno
 	}

container/landlock_test.go

@@ -0,0 +1,65 @@
+package container_test
+
+import (
+	"testing"
+	"unsafe"
+
+	"hakurei.app/container"
+)
+
+func TestLandlockString(t *testing.T) {
+	t.Parallel()
+
+	testCases := []struct {
+		name        string
+		rulesetAttr *container.RulesetAttr
+		want        string
+	}{
+		{"nil", nil, "NULL"},
+		{"zero", new(container.RulesetAttr), "0"},
+		{"some", &container.RulesetAttr{Scoped: container.LANDLOCK_SCOPE_SIGNAL}, "scoped: signal"},
+		{"set", &container.RulesetAttr{
+			HandledAccessFS:  container.LANDLOCK_ACCESS_FS_MAKE_SYM | container.LANDLOCK_ACCESS_FS_IOCTL_DEV | container.LANDLOCK_ACCESS_FS_WRITE_FILE,
+			HandledAccessNet: container.LANDLOCK_ACCESS_NET_BIND_TCP,
+			Scoped:           container.LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET | container.LANDLOCK_SCOPE_SIGNAL,
+		}, "fs: write_file make_sym fs_ioctl_dev, net: bind_tcp, scoped: abstract_unix_socket signal"},
+		{"all", &container.RulesetAttr{
+			HandledAccessFS: container.LANDLOCK_ACCESS_FS_EXECUTE |
+				container.LANDLOCK_ACCESS_FS_WRITE_FILE |
+				container.LANDLOCK_ACCESS_FS_READ_FILE |
+				container.LANDLOCK_ACCESS_FS_READ_DIR |
+				container.LANDLOCK_ACCESS_FS_REMOVE_DIR |
+				container.LANDLOCK_ACCESS_FS_REMOVE_FILE |
+				container.LANDLOCK_ACCESS_FS_MAKE_CHAR |
+				container.LANDLOCK_ACCESS_FS_MAKE_DIR |
+				container.LANDLOCK_ACCESS_FS_MAKE_REG |
+				container.LANDLOCK_ACCESS_FS_MAKE_SOCK |
+				container.LANDLOCK_ACCESS_FS_MAKE_FIFO |
+				container.LANDLOCK_ACCESS_FS_MAKE_BLOCK |
+				container.LANDLOCK_ACCESS_FS_MAKE_SYM |
+				container.LANDLOCK_ACCESS_FS_REFER |
+				container.LANDLOCK_ACCESS_FS_TRUNCATE |
+				container.LANDLOCK_ACCESS_FS_IOCTL_DEV,
+			HandledAccessNet: container.LANDLOCK_ACCESS_NET_BIND_TCP |
+				container.LANDLOCK_ACCESS_NET_CONNECT_TCP,
+			Scoped: container.LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET |
+				container.LANDLOCK_SCOPE_SIGNAL,
+		}, "fs: execute write_file read_file read_dir remove_dir remove_file make_char make_dir make_reg make_sock make_fifo make_block make_sym fs_refer fs_truncate fs_ioctl_dev, net: bind_tcp connect_tcp, scoped: abstract_unix_socket signal"},
+	}
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Parallel()
+			if got := tc.rulesetAttr.String(); got != tc.want {
+				t.Errorf("String: %s, want %s", got, tc.want)
+			}
+		})
+	}
+}
+
+func TestLandlockAttrSize(t *testing.T) {
+	t.Parallel()
+	want := 24
+	if got := unsafe.Sizeof(container.RulesetAttr{}); got != uintptr(want) {
+		t.Errorf("Sizeof: %d, want %d", got, want)
+	}
+}

container/mount.go

@@ -6,9 +6,8 @@ import (
 	"os"
 	. "syscall"
 
-	"hakurei.app/ext"
+	"hakurei.app/container/vfs"
 	"hakurei.app/message"
-	"hakurei.app/vfs"
 )
 
 /*
@@ -116,7 +115,7 @@ func (p *procPaths) remount(msg message.Msg, target string, flags uintptr) error
 	var targetKFinal string
 	{
 		var destFd int
-		if err := ext.IgnoringEINTR(func() (err error) {
+		if err := IgnoringEINTR(func() (err error) {
 			destFd, err = p.k.open(targetFinal, O_PATH|O_CLOEXEC, 0)
 			return
 		}); err != nil {

container/mount_test.go

@@ -5,8 +5,8 @@ import (
 	"syscall"
 	"testing"
 
-	"hakurei.app/internal/stub"
+	"hakurei.app/container/stub"
-	"hakurei.app/vfs"
+	"hakurei.app/container/vfs"
 )
 
 func TestBindMount(t *testing.T) {

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)
+			}
+		})
+	}
+}

container/params.go

@@ -0,0 +1,47 @@
+package container
+
+import (
+	"encoding/gob"
+	"errors"
+	"os"
+	"strconv"
+	"syscall"
+)
+
+// Setup appends the read end of a pipe for setup params transmission and returns its fd.
+func Setup(extraFiles *[]*os.File) (int, *os.File, error) {
+	if r, w, err := os.Pipe(); err != nil {
+		return -1, nil, err
+	} else {
+		fd := 3 + len(*extraFiles)
+		*extraFiles = append(*extraFiles, r)
+		return fd, w, nil
+	}
+}
+
+var (
+	ErrReceiveEnv = errors.New("environment variable not set")
+)
+
+// Receive retrieves setup fd from the environment and receives params.
+func Receive(key string, e any, fdp *uintptr) (func() error, error) {
+	var setup *os.File
+
+	if s, ok := os.LookupEnv(key); !ok {
+		return nil, ErrReceiveEnv
+	} else {
+		if fd, err := strconv.Atoi(s); err != nil {
+			return nil, optionalErrorUnwrap(err)
+		} else {
+			setup = os.NewFile(uintptr(fd), "setup")
+			if setup == nil {
+				return nil, syscall.EDOM
+			}
+			if fdp != nil {
+				*fdp = setup.Fd()
+			}
+		}
+	}
+
+	return setup.Close, gob.NewDecoder(setup).Decode(e)
+}

container/params_test.go

@@ -1,4 +1,4 @@
-package params_test
+package container_test
 
 import (
 	"encoding/gob"
@@ -9,7 +9,7 @@ import (
 	"syscall"
 	"testing"
 
-	"hakurei.app/internal/params"
+	"hakurei.app/container"
 )
 
 func TestSetupReceive(t *testing.T) {
@@ -30,8 +30,8 @@ func TestSetupReceive(t *testing.T) {
 			})
 		}
 
-		if _, err := params.Receive(key, nil, nil); !errors.Is(err, params.ErrReceiveEnv) {
+		if _, err := container.Receive(key, nil, nil); !errors.Is(err, container.ErrReceiveEnv) {
-			t.Errorf("Receive: error = %v, want %v", err, params.ErrReceiveEnv)
+			t.Errorf("Receive: error = %v, want %v", err, container.ErrReceiveEnv)
 		}
 	})
 
@@ -39,7 +39,7 @@ func TestSetupReceive(t *testing.T) {
 		const key = "TEST_ENV_FORMAT"
 		t.Setenv(key, "")
 
-		if _, err := params.Receive(key, nil, nil); !errors.Is(err, strconv.ErrSyntax) {
+		if _, err := container.Receive(key, nil, nil); !errors.Is(err, strconv.ErrSyntax) {
 			t.Errorf("Receive: error = %v, want %v", err, strconv.ErrSyntax)
 		}
 	})
@@ -48,7 +48,7 @@ func TestSetupReceive(t *testing.T) {
 		const key = "TEST_ENV_RANGE"
 		t.Setenv(key, "-1")
 
-		if _, err := params.Receive(key, nil, nil); !errors.Is(err, syscall.EDOM) {
+		if _, err := container.Receive(key, nil, nil); !errors.Is(err, syscall.EDOM) {
 			t.Errorf("Receive: error = %v, want %v", err, syscall.EDOM)
 		}
 	})
@@ -60,22 +60,16 @@ func TestSetupReceive(t *testing.T) {
 
 			encoderDone := make(chan error, 1)
 			extraFiles := make([]*os.File, 0, 1)
-			if r, w, err := os.Pipe(); err != nil {
+			deadline, _ := t.Deadline()
+			if fd, f, err := container.Setup(&extraFiles); err != nil {
 				t.Fatalf("Setup: error = %v", err)
+			} else if fd != 3 {
+				t.Fatalf("Setup: fd = %d, want 3", fd)
 			} else {
-				t.Cleanup(func() {
+				if err = f.SetDeadline(deadline); err != nil {
-					if err = errors.Join(r.Close(), w.Close()); err != nil {
+					t.Fatal(err.Error())
-						t.Fatal(err)
-					}
-				})
-
-				extraFiles = append(extraFiles, r)
-				if deadline, ok := t.Deadline(); ok {
-					if err = w.SetDeadline(deadline); err != nil {
-						t.Fatal(err)
-					}
 				}
-				go func() { encoderDone <- gob.NewEncoder(w).Encode(payload) }()
+				go func() { encoderDone <- gob.NewEncoder(f).Encode(payload) }()
 			}
 
 			if len(extraFiles) != 1 {
@@ -93,13 +87,13 @@ func TestSetupReceive(t *testing.T) {
 
 			var (
 				gotPayload []uint64
-				fdp        *int
+				fdp        *uintptr
 			)
 			if !useNilFdp {
-				fdp = new(int)
+				fdp = new(uintptr)
 			}
 			var closeFile func() error
-			if f, err := params.Receive(key, &gotPayload, fdp); err != nil {
+			if f, err := container.Receive(key, &gotPayload, fdp); err != nil {
 				t.Fatalf("Receive: error = %v", err)
 			} else {
 				closeFile = f
@@ -109,7 +103,7 @@ func TestSetupReceive(t *testing.T) {
 				}
 			}
 			if !useNilFdp {
-				if *fdp != dupFd {
+				if int(*fdp) != dupFd {
 					t.Errorf("Fd: %d, want %d", *fdp, dupFd)
 				}
 			}

container/path.go

@@ -4,21 +4,18 @@ import (
 	"errors"
 	"io/fs"
 	"os"
-	"path/filepath"
+	"path"
 	"strconv"
 	"strings"
 	"syscall"
 
-	"hakurei.app/fhs"
+	"hakurei.app/container/fhs"
-	"hakurei.app/vfs"
+	"hakurei.app/container/vfs"
 )
 
 const (
 	// Nonexistent is a path that cannot exist.
-	//
+	// /proc is chosen because a system with covered /proc is unsupported by this package.
-	// This path can never be presented by the kernel if proc is mounted on
-	// /proc/. This can only exist if parts of /proc/ is covered, or proc is not
-	// mounted at all. Neither configuration is supported by this package.
 	Nonexistent = fhs.Proc + "nonexistent"
 
 	hostPath    = fhs.Root + hostDir
@@ -29,16 +26,16 @@ const (
 
 func toSysroot(name string) string {
 	name = strings.TrimLeftFunc(name, func(r rune) bool { return r == '/' })
-	return filepath.Join(sysrootPath, name)
+	return path.Join(sysrootPath, name)
 }
 
 func toHost(name string) string {
 	name = strings.TrimLeftFunc(name, func(r rune) bool { return r == '/' })
-	return filepath.Join(hostPath, name)
+	return path.Join(hostPath, name)
 }
 
 func createFile(name string, perm, pperm os.FileMode, content []byte) error {
-	if err := os.MkdirAll(filepath.Dir(name), pperm); err != nil {
+	if err := os.MkdirAll(path.Dir(name), pperm); err != nil {
 		return err
 	}
 	f, err := os.OpenFile(name, syscall.O_CREAT|syscall.O_EXCL|syscall.O_WRONLY, perm)

container/path_test.go

@@ -4,14 +4,14 @@ import (
 	"io"
 	"math"
 	"os"
-	"path/filepath"
+	"path"
 	"reflect"
 	"syscall"
 	"testing"
 	"unsafe"
 
-	"hakurei.app/check"
+	"hakurei.app/container/check"
-	"hakurei.app/vfs"
+	"hakurei.app/container/vfs"
 )
 
 func TestToSysroot(t *testing.T) {
@@ -61,7 +61,7 @@ func TestCreateFile(t *testing.T) {
 				Path: "/proc/nonexistent",
 				Err:  syscall.ENOENT,
 			}
-			if err := createFile(filepath.Join(Nonexistent, ":3"), 0644, 0755, nil); !reflect.DeepEqual(err, wantErr) {
+			if err := createFile(path.Join(Nonexistent, ":3"), 0644, 0755, nil); !reflect.DeepEqual(err, wantErr) {
 				t.Errorf("createFile: error = %#v, want %#v", err, wantErr)
 			}
 		})
@@ -72,7 +72,7 @@ func TestCreateFile(t *testing.T) {
 				Path: "/proc/nonexistent",
 				Err:  syscall.ENOENT,
 			}
-			if err := createFile(filepath.Join(Nonexistent), 0644, 0755, nil); !reflect.DeepEqual(err, wantErr) {
+			if err := createFile(path.Join(Nonexistent), 0644, 0755, nil); !reflect.DeepEqual(err, wantErr) {
 				t.Errorf("createFile: error = %#v, want %#v", err, wantErr)
 			}
 		})
@@ -80,7 +80,7 @@ func TestCreateFile(t *testing.T) {
 
 	t.Run("touch", func(t *testing.T) {
 		tempDir := t.TempDir()
-		pathname := filepath.Join(tempDir, "empty")
+		pathname := path.Join(tempDir, "empty")
 		if err := createFile(pathname, 0644, 0755, nil); err != nil {
 			t.Fatalf("createFile: error = %v", err)
 		}
@@ -93,7 +93,7 @@ func TestCreateFile(t *testing.T) {
 
 	t.Run("write", func(t *testing.T) {
 		tempDir := t.TempDir()
-		pathname := filepath.Join(tempDir, "zero")
+		pathname := path.Join(tempDir, "zero")
 		if err := createFile(pathname, 0644, 0755, []byte{0}); err != nil {
 			t.Fatalf("createFile: error = %v", err)
 		}
@@ -107,7 +107,7 @@ func TestCreateFile(t *testing.T) {
 
 func TestEnsureFile(t *testing.T) {
 	t.Run("create", func(t *testing.T) {
-		if err := ensureFile(filepath.Join(t.TempDir(), "ensure"), 0644, 0755); err != nil {
+		if err := ensureFile(path.Join(t.TempDir(), "ensure"), 0644, 0755); err != nil {
 			t.Errorf("ensureFile: error = %v", err)
 		}
 	})
@@ -115,7 +115,7 @@ func TestEnsureFile(t *testing.T) {
 	t.Run("stat", func(t *testing.T) {
 		t.Run("inaccessible", func(t *testing.T) {
 			tempDir := t.TempDir()
-			pathname := filepath.Join(tempDir, "inaccessible")
+			pathname := path.Join(tempDir, "inaccessible")
 			if f, err := os.Create(pathname); err != nil {
 				t.Fatalf("Create: error = %v", err)
 			} else {
@@ -150,7 +150,7 @@ func TestEnsureFile(t *testing.T) {
 
 		t.Run("ensure", func(t *testing.T) {
 			tempDir := t.TempDir()
-			pathname := filepath.Join(tempDir, "ensure")
+			pathname := path.Join(tempDir, "ensure")
 			if f, err := os.Create(pathname); err != nil {
 				t.Fatalf("Create: error = %v", err)
 			} else {
@@ -195,12 +195,12 @@ func TestProcPaths(t *testing.T) {
 
 		t.Run("sample", func(t *testing.T) {
 			tempDir := t.TempDir()
-			if err := os.MkdirAll(filepath.Join(tempDir, "proc/self"), 0755); err != nil {
+			if err := os.MkdirAll(path.Join(tempDir, "proc/self"), 0755); err != nil {
 				t.Fatalf("MkdirAll: error = %v", err)
 			}
 
 			t.Run("clean", func(t *testing.T) {
-				if err := os.WriteFile(filepath.Join(tempDir, "proc/self/mountinfo"), []byte(`15 20 0:3 / /proc rw,relatime - proc /proc rw
+				if err := os.WriteFile(path.Join(tempDir, "proc/self/mountinfo"), []byte(`15 20 0:3 / /proc rw,relatime - proc /proc rw
 16 20 0:15 / /sys rw,relatime - sysfs /sys rw
 17 20 0:5 / /dev rw,relatime - devtmpfs udev rw,size=1983516k,nr_inodes=495879,mode=755`), 0644); err != nil {
 					t.Fatalf("WriteFile: error = %v", err)
@@ -243,8 +243,8 @@ func TestProcPaths(t *testing.T) {
 			})
 
 			t.Run("malformed", func(t *testing.T) {
-				filepath.Join(tempDir, "proc/self/mountinfo")
+				path.Join(tempDir, "proc/self/mountinfo")
-				if err := os.WriteFile(filepath.Join(tempDir, "proc/self/mountinfo"), []byte{0}, 0644); err != nil {
+				if err := os.WriteFile(path.Join(tempDir, "proc/self/mountinfo"), []byte{0}, 0644); err != nil {
 					t.Fatalf("WriteFile: error = %v", err)
 				}
 

container/seccomp/libseccomp.go

@@ -16,7 +16,6 @@ import (
 	"unsafe"
 
 	"hakurei.app/container/std"
-	"hakurei.app/ext"
 )
 
 // ErrInvalidRules is returned for a zero-length rules slice.
@@ -89,22 +88,18 @@ var resPrefix = [...]string{
 	7: "seccomp_load failed",
 }
 
-// cbAllocateBuffer is the function signature for the function handle passed to
+// cbAllocateBuffer is the function signature for the function handle passed to hakurei_export_filter
-// hakurei_scmp_make_filter which allocates the buffer that the resulting bpf
+// which allocates the buffer that the resulting bpf program is copied into, and writes its slice header
-// program is copied into, and writes its slice header to a value held by the caller.
+// to a value held by the caller.
 type cbAllocateBuffer = func(len C.size_t) (buf unsafe.Pointer)
 
-// hakurei_scmp_allocate allocates a buffer of specified size known to the
-// runtime through a callback passed in a [cgo.Handle].
-//
 //export hakurei_scmp_allocate
 func hakurei_scmp_allocate(f C.uintptr_t, len C.size_t) (buf unsafe.Pointer) {
 	return cgo.Handle(f).Value().(cbAllocateBuffer)(len)
 }
 
-// makeFilter generates a bpf program from a slice of [std.NativeRule] and
+// makeFilter generates a bpf program from a slice of [std.NativeRule] and writes the resulting byte slice to p.
-// writes the resulting byte slice to p. The filter is installed to the current
+// The filter is installed to the current process if p is nil.
-// process if p is nil.
 func makeFilter(rules []std.NativeRule, flags ExportFlag, p *[]byte) error {
 	if len(rules) == 0 {
 		return ErrInvalidRules
@@ -175,8 +170,8 @@ func Export(rules []std.NativeRule, flags ExportFlag) (data []byte, err error) {
 	return
 }
 
-// Load generates a bpf program from a slice of [std.NativeRule] and enforces it
+// Load generates a bpf program from a slice of [std.NativeRule] and enforces it on the current process.
-// on the current process. Errors returned by libseccomp is wrapped in [LibraryError].
+// Errors returned by libseccomp is wrapped in [LibraryError].
 func Load(rules []std.NativeRule, flags ExportFlag) error { return makeFilter(rules, flags, nil) }
 
 type (
@@ -220,9 +215,9 @@ const (
 
 // syscallResolveName resolves a syscall number by name via seccomp_syscall_resolve_name.
 // This function is only for testing the lookup tables and included here for convenience.
-func syscallResolveName(s string) (num ext.SyscallNum, ok bool) {
+func syscallResolveName(s string) (num std.ScmpSyscall, ok bool) {
 	v := C.CString(s)
-	num = ext.SyscallNum(C.seccomp_syscall_resolve_name(v))
+	num = std.ScmpSyscall(C.seccomp_syscall_resolve_name(v))
 	C.free(unsafe.Pointer(v))
 	ok = num != C.__NR_SCMP_ERROR
 	return

container/seccomp/presets.go

@@ -6,7 +6,6 @@ import (
 	. "syscall"
 
 	. "hakurei.app/container/std"
-	. "hakurei.app/ext"
 )
 
 func Preset(presets FilterPreset, flags ExportFlag) (rules []NativeRule) {

container/seccomp/std_test.go

@@ -6,13 +6,12 @@ import (
 	"unsafe"
 
 	"hakurei.app/container/std"
-	"hakurei.app/ext"
 )
 
 func TestSyscallResolveName(t *testing.T) {
 	t.Parallel()
 
-	for name, want := range ext.Syscalls() {
+	for name, want := range std.Syscalls() {
 		t.Run(name, func(t *testing.T) {
 			t.Parallel()
 
@@ -25,10 +24,8 @@ func TestSyscallResolveName(t *testing.T) {
 }
 
 func TestRuleType(t *testing.T) {
-	assertKind[ext.Uint, scmpUint](t)
+	assertKind[std.ScmpUint, scmpUint](t)
-	assertOverflow(t, ext.Uint(ext.MaxUint))
+	assertKind[std.ScmpInt, scmpInt](t)
-	assertKind[ext.Int, scmpInt](t)
-	assertOverflow(t, ext.Int(ext.MaxInt))
 
 	assertSize[std.NativeRule, syscallRule](t)
 	assertKind[std.ScmpDatum, scmpDatum](t)
@@ -64,14 +61,3 @@ func assertKind[native, equivalent any](t *testing.T) {
 		t.Fatalf("%s: %s, want %s", nativeType.Name(), nativeType.Kind(), equivalentType.Kind())
 	}
 }
-
-// assertOverflow asserts that incrementing m overflows.
-func assertOverflow[T ~int32 | ~uint32](t *testing.T, m T) {
-	t.Helper()
-
-	old := m
-	m++
-	if m > old {
-		t.Fatalf("unexpected value %#x", m)
-	}
-}

container/std/mksysnum_linux.pl

@@ -19,11 +19,11 @@ print <<EOF;
 // $command
 // Code generated by the command above; DO NOT EDIT.
 
-package ext
+package std
 
 import . "syscall"
 
-var syscallNum = map[string]SyscallNum{
+var syscallNum = map[string]ScmpSyscall{
 EOF
 
 my $offset = 0;
@@ -45,7 +45,7 @@ sub fmt {
 		print "	\"$name\": SNR_$name_upper,\n";
 	}
 	elsif($state == 1){
-		print " SNR_$name_upper SyscallNum = SYS_$name_upper\n";
+		print " SNR_$name_upper ScmpSyscall = SYS_$name_upper\n";
 	}
 	else{
 		return;

container/std/pnr.go

@@ -1,6 +1,6 @@
 // Code generated from include/seccomp-syscalls.h; DO NOT EDIT.
 
-package ext
+package std
 
 /*
  * pseudo syscall definitions

container/std/seccomp.go

@@ -1,20 +1,30 @@
 package std
 
-import "hakurei.app/ext"
+import (
+	"encoding/json"
+	"strconv"
+)
 
 type (
+	// ScmpUint is equivalent to C.uint.
+	ScmpUint uint32
+	// ScmpInt is equivalent to C.int.
+	ScmpInt int32
+
+	// ScmpSyscall represents a syscall number passed to libseccomp via [NativeRule.Syscall].
+	ScmpSyscall ScmpInt
 	// ScmpErrno represents an errno value passed to libseccomp via [NativeRule.Errno].
-	ScmpErrno = ext.Int
+	ScmpErrno ScmpInt
 
 	// ScmpCompare is equivalent to enum scmp_compare;
-	ScmpCompare = ext.Uint
+	ScmpCompare ScmpUint
 	// ScmpDatum is equivalent to scmp_datum_t.
-	ScmpDatum = uint64
+	ScmpDatum uint64
 
 	// ScmpArgCmp is equivalent to struct scmp_arg_cmp.
 	ScmpArgCmp struct {
 		// argument number, starting at 0
-		Arg ext.Uint `json:"arg"`
+		Arg ScmpUint `json:"arg"`
 		// the comparison op, e.g. SCMP_CMP_*
 		Op ScmpCompare `json:"op"`
 
@@ -25,10 +35,42 @@ type (
 	// A NativeRule specifies an arch-specific action taken by seccomp under certain conditions.
 	NativeRule struct {
 		// Syscall is the arch-dependent syscall number to act against.
-		Syscall ext.SyscallNum `json:"syscall"`
+		Syscall ScmpSyscall `json:"syscall"`
 		// Errno is the errno value to return when the condition is satisfied.
 		Errno ScmpErrno `json:"errno"`
 		// Arg is the optional struct scmp_arg_cmp passed to libseccomp.
 		Arg *ScmpArgCmp `json:"arg,omitempty"`
 	}
 )
+
+// MarshalJSON resolves the name of [ScmpSyscall] and encodes it as a [json] string.
+// If such a name does not exist, the syscall number is encoded instead.
+func (num *ScmpSyscall) MarshalJSON() ([]byte, error) {
+	n := *num
+	for name, cur := range Syscalls() {
+		if cur == n {
+			return json.Marshal(name)
+		}
+	}
+	return json.Marshal(n)
+}
+
+// SyscallNameError is returned when trying to unmarshal an invalid syscall name into [ScmpSyscall].
+type SyscallNameError string
+
+func (e SyscallNameError) Error() string { return "invalid syscall name " + strconv.Quote(string(e)) }
+
+// UnmarshalJSON looks up the syscall number corresponding to name encoded in data
+// by calling [SyscallResolveName].
+func (num *ScmpSyscall) UnmarshalJSON(data []byte) error {
+	var name string
+	if err := json.Unmarshal(data, &name); err != nil {
+		return err
+	}
+	if n, ok := SyscallResolveName(name); !ok {
+		return SyscallNameError(name)
+	} else {
+		*num = n
+		return nil
+	}
+}

container/std/seccomp_test.go

@@ -1,4 +1,4 @@
-package ext_test
+package std_test
 
 import (
 	"encoding/json"
@@ -7,39 +7,39 @@ import (
 	"reflect"
 	"testing"
 
-	"hakurei.app/ext"
+	"hakurei.app/container/std"
 )
 
-func TestSyscall(t *testing.T) {
+func TestScmpSyscall(t *testing.T) {
 	t.Parallel()
 
 	testCases := []struct {
 		name string
 		data string
-		want ext.SyscallNum
+		want std.ScmpSyscall
 		err  error
 	}{
-		{"epoll_create1", `"epoll_create1"`, ext.SNR_EPOLL_CREATE1, nil},
+		{"epoll_create1", `"epoll_create1"`, std.SNR_EPOLL_CREATE1, nil},
-		{"clone3", `"clone3"`, ext.SNR_CLONE3, nil},
+		{"clone3", `"clone3"`, std.SNR_CLONE3, nil},
 
 		{"oob", `-2147483647`, -math.MaxInt32,
 			&json.UnmarshalTypeError{Value: "number", Type: reflect.TypeFor[string](), Offset: 11}},
 		{"name", `"nonexistent_syscall"`, -math.MaxInt32,
-			ext.SyscallNameError("nonexistent_syscall")},
+			std.SyscallNameError("nonexistent_syscall")},
 	}
 	for _, tc := range testCases {
 		t.Run(tc.name, func(t *testing.T) {
 			t.Parallel()
 
 			t.Run("decode", func(t *testing.T) {
-				var got ext.SyscallNum
+				var got std.ScmpSyscall
 				if err := json.Unmarshal([]byte(tc.data), &got); !reflect.DeepEqual(err, tc.err) {
 					t.Fatalf("Unmarshal: error = %#v, want %#v", err, tc.err)
 				} else if err == nil && got != tc.want {
 					t.Errorf("Unmarshal: %v, want %v", got, tc.want)
 				}
 			})
-			if errors.As(tc.err, new(ext.SyscallNameError)) {
+			if errors.As(tc.err, new(std.SyscallNameError)) {
 				return
 			}
 
@@ -55,22 +55,8 @@ func TestSyscall(t *testing.T) {
 
 	t.Run("error", func(t *testing.T) {
 		const want = `invalid syscall name "\x00"`
-		if got := ext.SyscallNameError("\x00").Error(); got != want {
+		if got := std.SyscallNameError("\x00").Error(); got != want {
 			t.Fatalf("Error: %q, want %q", got, want)
 		}
 	})
 }
-
-func TestSyscallResolveName(t *testing.T) {
-	t.Parallel()
-
-	for name, want := range ext.Syscalls() {
-		t.Run(name, func(t *testing.T) {
-			t.Parallel()
-
-			if got, ok := ext.SyscallResolveName(name); !ok || got != want {
-				t.Errorf("SyscallResolveName(%q) = %d, want %d", name, got, want)
-			}
-		})
-	}
-}

container/std/syscall.go

@@ -0,0 +1,28 @@
+package std
+
+import "iter"
+
+// Syscalls returns an iterator over all wired syscalls.
+func Syscalls() iter.Seq2[string, ScmpSyscall] {
+	return func(yield func(string, ScmpSyscall) bool) {
+		for name, num := range syscallNum {
+			if !yield(name, num) {
+				return
+			}
+		}
+		for name, num := range syscallNumExtra {
+			if !yield(name, num) {
+				return
+			}
+		}
+	}
+}
+
+// SyscallResolveName resolves a syscall number from its string representation.
+func SyscallResolveName(name string) (num ScmpSyscall, ok bool) {
+	if num, ok = syscallNum[name]; ok {
+		return
+	}
+	num, ok = syscallNumExtra[name]
+	return
+}

container/std/syscall_extra_linux_386.go

@@ -0,0 +1,13 @@
+package std
+
+var syscallNumExtra = map[string]ScmpSyscall{
+	"kexec_file_load": SNR_KEXEC_FILE_LOAD,
+	"subpage_prot":    SNR_SUBPAGE_PROT,
+	"switch_endian":   SNR_SWITCH_ENDIAN,
+}
+
+const (
+	SNR_KEXEC_FILE_LOAD ScmpSyscall = __PNR_kexec_file_load
+	SNR_SUBPAGE_PROT    ScmpSyscall = __PNR_subpage_prot
+	SNR_SWITCH_ENDIAN   ScmpSyscall = __PNR_switch_endian
+)