Re: [RFC PATCH 07/13] rootns: Add rootns_enter() system call
From: Andy Lutomirski
Date: Tue Nov 17 2026 - 15:54:07 EST
On Tue, Nov 17, 2026 at 12:03 PM Jarkko Sakkinen <jarkko@xxxxxxxxxx> wrote:
>
> On Tue, Nov 17, 2026 at 09:42:26AM -0800, Andy Lutomirski wrote:
> > On Mon, Nov 16, 2026 at 9:09 PM Jarkko Sakkinen <jarkko@xxxxxxxxxx> wrote:
> > >
> > > Multithreaded userspace daemons cannot call setns(CLONE_NEWUSER) or
> > > safely mutate process-wide filesystem state in place before spawning a
> > > child into a target namespace set.
> > >
> > > Add the rootns_enter() system call to fork a child process directly into
> > > a rootns whose root mount has been installed. Transition the rootns from
> > > ROOTNS_READY through ROOTNS_FORKING to ROOTNS_RUNNING for the first
> > > process so it becomes PID 1, serialize concurrent entrants on waitq
> > > until bootstrap resolves, and admit subsequent callers into the running
> > > PID namespace until teardown begins.
> > >
> > > Require CAP_SYS_ADMIN in the target user namespace, forbid credential or
> > > PR_SET_NO_NEW_PRIVS privilege escalation
> >
> > Why are you disallowing this if no_new_privs?
>
> For no any good reason at all apparently: no_new_privs is inherited by
> the child and only restricts privilege gain across execve(), whereas
> user_namespace transitions are allowed under it.
>
> So yeah, it's a spurious check, and thus I'll drop it.
>
> > This isn't necessarily a showstopper, but fork() is really quite slow
> > in many cases. It would be nice to have a more spawn-like API
> > available. I assume it's possible to open a file in the target's
> > mountns by some sane API even without entering that mountns, as a side
> > effect of how the fsopen / move_mount mechanism works.
>
>
> If the caller retains the O_PATH fd returned by fsmount() for "/", it is
> able to open files inside the target mount tree via openat2(...,
> RESOLVE_IN_ROOT) w/o entering the mount namespace, and pass the
> resulting fd to execveat().
>
> For avoiding full fork(), there are (at least) two options:
>
> 1. A rootns flag in clone3() (e.g., CLONE_INTO_ROOTNS) and a rootns fd
> in struct clone_args (disallowing incompatible CLONE_NEW*, CLONE_FS,
> CLONE_THREAD, ...), and callers can pair it with CLONE_VM | CLONE_VFORK.
> 2. A flags parameter in rootns_enter() (e.g., ROOTNS_ENTER_VFORK), which
> would map directly to CLONE_VM | CLONE_VFORK.
There's also been some work on using a pidfd-like object to build and
spawn a subprocess. Supporting that would make sense, but I don't
know to what extent these projects should block each other.
--
Andy Lutomirski
AMA Capital Management, LLC