app: integrate container autoroot

Doing this instead of mounting directly on / because it's impossible to ensure a parent is available for every path hakurei wants to mount to. This situation is similar to autoetc hence the similar name, however a symlink mirror will not work in this case.

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

hst/container.go

@@ -45,10 +45,17 @@ type (
 		// create symlinks inside container filesystem
 		Link [][2]string `json:"symlink"`
 
+		// automatically bind mount top-level directories to container root;
+		// the zero value disables this behaviour
+		AutoRoot string `json:"auto_root,omitempty"`
+		// extra flags for AutoRoot
+		RootFlags int `json:"root_flags,omitempty"`
+
 		// read-only /etc directory
 		Etc string `json:"etc,omitempty"`
 		// automatically set up /etc symlinks
 		AutoEtc bool `json:"auto_etc"`
+
 		// cover these paths or create them if they do not already exist
 		Cover []string `json:"cover"`
 	}

internal/app/app_pd_linux_test.go

@@ -42,21 +42,10 @@ var testCasesPd = []sealTestCase{
 				"XDG_SESSION_TYPE=tty",
 			},
 			Ops: new(container.Ops).
+				Root("/", "4a450b6596d7bc15bd01780eb9a607ac", container.BindWritable).
 				Proc("/proc").
 				Tmpfs(hst.Tmp, 4096, 0755).
 				Dev("/dev").Mqueue("/dev/mqueue").
-				Bind("/bin", "/bin", container.BindWritable).
-				Bind("/boot", "/boot", container.BindWritable).
-				Bind("/home", "/home", container.BindWritable).
-				Bind("/lib", "/lib", container.BindWritable).
-				Bind("/lib64", "/lib64", container.BindWritable).
-				Bind("/nix", "/nix", container.BindWritable).
-				Bind("/root", "/root", container.BindWritable).
-				Bind("/run", "/run", container.BindWritable).
-				Bind("/srv", "/srv", container.BindWritable).
-				Bind("/sys", "/sys", container.BindWritable).
-				Bind("/usr", "/usr", container.BindWritable).
-				Bind("/var", "/var", container.BindWritable).
 				Bind("/dev/kvm", "/dev/kvm", container.BindWritable|container.BindDevice|container.BindOptional).
 				Tmpfs("/run/user/1971", 8192, 0755).
 				Tmpfs("/run/dbus", 8192, 0755).
@@ -186,21 +175,10 @@ var testCasesPd = []sealTestCase{
 				"XDG_SESSION_TYPE=tty",
 			},
 			Ops: new(container.Ops).
+				Root("/", "ebf083d1b175911782d413369b64ce7c", container.BindWritable).
 				Proc("/proc").
 				Tmpfs(hst.Tmp, 4096, 0755).
 				Dev("/dev").Mqueue("/dev/mqueue").
-				Bind("/bin", "/bin", container.BindWritable).
-				Bind("/boot", "/boot", container.BindWritable).
-				Bind("/home", "/home", container.BindWritable).
-				Bind("/lib", "/lib", container.BindWritable).
-				Bind("/lib64", "/lib64", container.BindWritable).
-				Bind("/nix", "/nix", container.BindWritable).
-				Bind("/root", "/root", container.BindWritable).
-				Bind("/run", "/run", container.BindWritable).
-				Bind("/srv", "/srv", container.BindWritable).
-				Bind("/sys", "/sys", container.BindWritable).
-				Bind("/usr", "/usr", container.BindWritable).
-				Bind("/var", "/var", container.BindWritable).
 				Bind("/dev/dri", "/dev/dri", container.BindWritable|container.BindDevice|container.BindOptional).
 				Bind("/dev/kvm", "/dev/kvm", container.BindWritable|container.BindDevice|container.BindOptional).
 				Tmpfs("/run/user/1971", 8192, 0755).

internal/app/container_linux.go

@@ -6,6 +6,7 @@ import (
 	"io/fs"
 	"maps"
 	"path"
+	"slices"
 	"syscall"
 
 	"hakurei.app/container"
@@ -21,7 +22,7 @@ const preallocateOpsCount = 1 << 5
 
 // newContainer initialises [container.Params] via [hst.ContainerConfig].
 // Note that remaining container setup must be queued by the caller.
-func newContainer(s *hst.ContainerConfig, os sys.State, uid, gid *int) (*container.Params, map[string]string, error) {
+func newContainer(s *hst.ContainerConfig, os sys.State, prefix string, uid, gid *int) (*container.Params, map[string]string, error) {
 	if s == nil {
 		return nil, nil, syscall.EBADE
 	}
@@ -72,6 +73,13 @@ func newContainer(s *hst.ContainerConfig, os sys.State, uid, gid *int) (*contain
 		*gid = container.OverflowGid()
 	}
 
+	if s.AutoRoot != "" {
+		if !path.IsAbs(s.AutoRoot) {
+			return nil, nil, fmt.Errorf("auto root target %q not absolute", s.AutoRoot)
+		}
+		params.Root(s.AutoRoot, prefix, s.RootFlags)
+	}
+
 	params.
 		Proc("/proc").
 		Tmpfs(hst.Tmp, 1<<12, 0755)
@@ -120,6 +128,28 @@ func newContainer(s *hst.ContainerConfig, os sys.State, uid, gid *int) (*contain
 			return nil, nil, err
 		}
 	}
+	// evaluated path, input path
+	hidePathSource := make([][2]string, 0, len(s.Filesystem))
+
+	// AutoRoot is a collection of many BindMountOp internally
+	if s.AutoRoot != "" {
+		if d, err := os.ReadDir(s.AutoRoot); err != nil {
+			return nil, nil, err
+		} else {
+			hidePathSource = slices.Grow(hidePathSource, len(d))
+			for _, ent := range d {
+				name := ent.Name()
+				if container.IsAutoRootBindable(name) {
+					name = path.Join(s.AutoRoot, name)
+					srcP := [2]string{name, name}
+					if err = evalSymlinks(os, &srcP[0]); err != nil {
+						return nil, nil, err
+					}
+					hidePathSource = append(hidePathSource, srcP)
+				}
+			}
+		}
+	}
 
 	for _, c := range s.Filesystem {
 		if c == nil {
@@ -137,24 +167,11 @@ func newContainer(s *hst.ContainerConfig, os sys.State, uid, gid *int) (*contain
 			return nil, nil, fmt.Errorf("dst path %q is not absolute", dest)
 		}
 
-		srcH := c.Src
+		p := [2]string{c.Src, c.Src}
-		if err := evalSymlinks(os, &srcH); err != nil {
+		if err := evalSymlinks(os, &p[0]); err != nil {
 			return nil, nil, err
 		}
-
+		hidePathSource = append(hidePathSource, p)
-		for i := range hidePaths {
-			// skip matched entries
-			if hidePathMatch[i] {
-				continue
-			}
-
-			if ok, err := deepContainsH(srcH, hidePaths[i]); err != nil {
-				return nil, nil, err
-			} else if ok {
-				hidePathMatch[i] = true
-				os.Printf("hiding paths from %q", c.Src)
-			}
-		}
 
 		var flags int
 		if c.Write {
@@ -169,6 +186,22 @@ func newContainer(s *hst.ContainerConfig, os sys.State, uid, gid *int) (*contain
 		params.Bind(c.Src, dest, flags)
 	}
 
+	for _, p := range hidePathSource {
+		for i := range hidePaths {
+			// skip matched entries
+			if hidePathMatch[i] {
+				continue
+			}
+
+			if ok, err := deepContainsH(p[0], hidePaths[i]); err != nil {
+				return nil, nil, err
+			} else if ok {
+				hidePathMatch[i] = true
+				os.Printf("hiding path %q from %q", hidePaths[i], p[1])
+			}
+		}
+	}
+
 	// cover matched paths
 	for i, ok := range hidePathMatch {
 		if ok {
@@ -180,6 +213,18 @@ func newContainer(s *hst.ContainerConfig, os sys.State, uid, gid *int) (*contain
 		params.Link(l[0], l[1])
 	}
 
+	if !s.AutoEtc {
+		if s.Etc != "" {
+			params.Bind(s.Etc, "/etc", 0)
+		}
+	} else {
+		etcPath := s.Etc
+		if etcPath == "" {
+			etcPath = "/etc"
+		}
+		params.Etc(etcPath, prefix)
+	}
+
 	return params, maps.Clone(s.Env), nil
 }
 

internal/app/seal_linux.go

@@ -241,26 +241,9 @@ func (seal *outcome) finalise(ctx context.Context, sys sys.State, config *hst.Co
 			Net:     true,
 			Tty:     true,
 			AutoEtc: true,
-		}
-		// bind entries in /
-		if d, err := sys.ReadDir("/"); err != nil {
-			return err
-		} else {
-			b := make([]*hst.FilesystemConfig, 0, len(d))
-			for _, ent := range d {
-				p := "/" + ent.Name()
-				switch p {
-				case "/proc":
-				case "/dev":
-				case "/tmp":
-				case "/mnt":
-				case "/etc":
 
-				default:
+			AutoRoot:  "/",
-					b = append(b, &hst.FilesystemConfig{Src: p, Write: true, Must: true})
+			RootFlags: container.BindWritable,
-				}
-			}
-			conf.Filesystem = append(conf.Filesystem, b...)
 		}
 
 		// hide nscd from sandbox if present
@@ -282,7 +265,7 @@ func (seal *outcome) finalise(ctx context.Context, sys sys.State, config *hst.Co
 	{
 		var uid, gid int
 		var err error
-		seal.container, seal.env, err = newContainer(config.Container, sys, &uid, &gid)
+		seal.container, seal.env, err = newContainer(config.Container, sys, seal.id.String(), &uid, &gid)
 		seal.waitDelay = config.Container.WaitDelay
 		if err != nil {
 			return hlog.WrapErrSuffix(err,
@@ -305,18 +288,6 @@ func (seal *outcome) finalise(ctx context.Context, sys sys.State, config *hst.Co
 		}
 	}
 
-	if !config.Container.AutoEtc {
-		if config.Container.Etc != "" {
-			seal.container.Bind(config.Container.Etc, "/etc", 0)
-		}
-	} else {
-		etcPath := config.Container.Etc
-		if etcPath == "" {
-			etcPath = "/etc"
-		}
-		seal.container.Etc(etcPath, seal.id.String())
-	}
-
 	// inner XDG_RUNTIME_DIR default formatting of `/run/user/%d` as mapped uid
 	innerRuntimeDir := path.Join("/run/user", mapuid.String())
 	seal.env[xdgRuntimeDir] = innerRuntimeDir

test/sandbox/case/pd.nix

@@ -139,18 +139,6 @@
 
     mount = [
       (ent "/sysroot" "/" "rw,nosuid,nodev,relatime" "tmpfs" "rootfs" "rw,uid=1000000,gid=1000000")
-      (ent "/" "/proc" "rw,nosuid,nodev,noexec,relatime" "proc" "proc" "rw")
-      (ent "/" "/.hakurei" "rw,nosuid,nodev,relatime" "tmpfs" "tmpfs" "rw,size=4k,mode=755,uid=1000000,gid=1000000")
-      (ent "/" "/dev" "rw,nosuid,nodev,relatime" "tmpfs" "devtmpfs" "rw,mode=755,uid=1000000,gid=1000000")
-      (ent "/null" "/dev/null" "rw,nosuid" "devtmpfs" "devtmpfs" ignore)
-      (ent "/zero" "/dev/zero" "rw,nosuid" "devtmpfs" "devtmpfs" ignore)
-      (ent "/full" "/dev/full" "rw,nosuid" "devtmpfs" "devtmpfs" ignore)
-      (ent "/random" "/dev/random" "rw,nosuid" "devtmpfs" "devtmpfs" ignore)
-      (ent "/urandom" "/dev/urandom" "rw,nosuid" "devtmpfs" "devtmpfs" ignore)
-      (ent "/tty" "/dev/tty" "rw,nosuid" "devtmpfs" "devtmpfs" ignore)
-      (ent "/" "/dev/pts" "rw,nosuid,noexec,relatime" "devpts" "devpts" "rw,mode=620,ptmxmode=666")
-      (ent ignore "/dev/console" "rw,nosuid,noexec,relatime" "devpts" "devpts" "rw,gid=3,mode=620,ptmxmode=666")
-      (ent "/" "/dev/mqueue" "rw,nosuid,nodev,noexec,relatime" "mqueue" "mqueue" "rw")
       (ent "/bin" "/bin" "rw,nosuid,nodev,relatime" "ext4" "/dev/disk/by-label/nixos" "rw")
       (ent "/home" "/home" "rw,nosuid,nodev,relatime" "ext4" "/dev/disk/by-label/nixos" "rw")
       (ent "/lib64" "/lib64" "rw,nosuid,nodev,relatime" "ext4" "/dev/disk/by-label/nixos" "rw")
@@ -180,6 +168,18 @@
       (ent "/" ignore "rw,nosuid,nodev,noexec,relatime" ignore ignore "rw")
       (ent "/usr" "/usr" "rw,nosuid,nodev,relatime" "ext4" "/dev/disk/by-label/nixos" "rw")
       (ent "/var" "/var" "rw,nosuid,nodev,relatime" "ext4" "/dev/disk/by-label/nixos" "rw")
+      (ent "/" "/proc" "rw,nosuid,nodev,noexec,relatime" "proc" "proc" "rw")
+      (ent "/" "/.hakurei" "rw,nosuid,nodev,relatime" "tmpfs" "tmpfs" "rw,size=4k,mode=755,uid=1000000,gid=1000000")
+      (ent "/" "/dev" "rw,nosuid,nodev,relatime" "tmpfs" "devtmpfs" "rw,mode=755,uid=1000000,gid=1000000")
+      (ent "/null" "/dev/null" "rw,nosuid" "devtmpfs" "devtmpfs" ignore)
+      (ent "/zero" "/dev/zero" "rw,nosuid" "devtmpfs" "devtmpfs" ignore)
+      (ent "/full" "/dev/full" "rw,nosuid" "devtmpfs" "devtmpfs" ignore)
+      (ent "/random" "/dev/random" "rw,nosuid" "devtmpfs" "devtmpfs" ignore)
+      (ent "/urandom" "/dev/urandom" "rw,nosuid" "devtmpfs" "devtmpfs" ignore)
+      (ent "/tty" "/dev/tty" "rw,nosuid" "devtmpfs" "devtmpfs" ignore)
+      (ent "/" "/dev/pts" "rw,nosuid,noexec,relatime" "devpts" "devpts" "rw,mode=620,ptmxmode=666")
+      (ent ignore "/dev/console" "rw,nosuid,noexec,relatime" "devpts" "devpts" "rw,gid=3,mode=620,ptmxmode=666")
+      (ent "/" "/dev/mqueue" "rw,nosuid,nodev,noexec,relatime" "mqueue" "mqueue" "rw")
       (ent "/kvm" "/dev/kvm" "rw,nosuid" "devtmpfs" "devtmpfs" ignore)
       (ent "/" "/run/user/1000" "rw,nosuid,nodev,relatime" "tmpfs" "tmpfs" "rw,size=8k,mode=755,uid=1000000,gid=1000000")
       (ent "/" "/run/dbus" "rw,nosuid,nodev,relatime" "tmpfs" "tmpfs" "rw,size=8k,mode=755,uid=1000000,gid=1000000")