container: optionally map uid/gid 0 as init

Unfortunately required to work around flawed APIs like binfmt_misc.

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

container/container.go

@@ -91,6 +91,9 @@ type (
 		// Time to wait for processes lingering after the initial process terminates.
 		AdoptWaitDelay time.Duration
 
+		// Map uid/gid 0 in the init process. Requires [FstypeProc] attached to
+		// [fhs.Proc] in the container filesystem.
+		InitAsRoot bool
 		// Mapped Uid in user namespace.
 		Uid int
 		// Mapped Gid in user namespace.
@@ -286,6 +289,18 @@ func (p *Container) Start() error {
 	if !p.HostNet {
 		p.cmd.SysProcAttr.Cloneflags |= CLONE_NEWNET
 	}
+	if p.InitAsRoot {
+		p.cmd.SysProcAttr.AmbientCaps = append(p.cmd.SysProcAttr.AmbientCaps,
+			// mappings during init as root
+			CAP_SETFCAP,
+		)
+
+		if !p.SeccompDisable &&
+			len(p.SeccompRules) == 0 &&
+			p.SeccompPresets&std.PresetDenyNS != 0 {
+			return errors.New("container: as root requires late namespace creation")
+		}
+	}
 
 	// place setup pipe before user supplied extra files, this is later restored by init
 	if r, w, err := os.Pipe(); err != nil {