Re: [PATCH] pasta: Add --no-pidns to keep spawned command in caller's PID namespace

[Stefano Brivio <sbrivio@redhat.com>]

On Tue, 8 Sep 2026 16:03:07 +1000
David Gibson <david@gibson.dropbear.id.au> wrote:

> On Sun, Sep 06, 2026 at 03:17:58PM +0200, Christian Korneck wrote:
> > This is to allow running pasta inside a container without unmasking
> > /proc for the whole container (Docker's --security-opt
> > systempaths=unconfined, Podman's --security-opt unmask=ALL), which is
> > undesirable as it exposes /proc/sysrq-trigger and other masked paths.
> > 
> > In spawn mode, pasta clones the command with CLONE_NEWPID and mounts a
> > new procfs instance on /proc, so that it matches the new PID namespace.
> > 
> > Mounting procfs in a new user namespace requires a fully visible,
> > unobstructed procfs. Container runtimes deliberately obstruct /proc
> > (Docker, for example, masks /proc/kcore and friends and mounts
> > /proc/sys read-only), so the mount is refused:
> > 
> >   Couldn't mount /proc: Operation not permitted
> > 
> > We only warn and continue, leaving the command in a new PID namespace
> > while the visible /proc still numbers processes in the outer one.
> > Anything resolving its own PID through /proc then fails, for example
> > bubblewrap:
> > 
> >   bwrap: open /proc/22/ns/ns failed: No such file or directory
> > 
> > Add a --no-pidns option: skip CLONE_NEWPID for the spawned command and
> > don't mount /proc, which is then not needed. User, network, mount, UTS
> > and IPC namespaces, --config-net and port forwarding are unaffected.
> > The option is rejected together with PID or --netns, as it only makes
> > sense when we spawn the command ourselves.
> > 
> > Add a test checking that, by default, the command runs in a new PID
> > namespace, and that --no-pidns keeps it in the caller's one.  
> 
> I can see that this would be useful.  On the other hand, I could say
> that for just about any combinatiom of unshare(1) options, and I don't
> think we want to reimplement all of unshare(1) in pasta.

Certainly we don't, but we don't get requests like these frequently
(except for the one to drop namespacing altogether, which is another
story), so I think it's acceptable.

The patch looks also pretty good and complete, so really low effort from
my side.

> Note that it
> is always possible to create a namespace using unshare(1) with
> arbitrary options and then connect it with pasta, rather than having
> pasta create its own.
> 
> That said, I've never seen the point of creating a pidns by default in
> pasta.  We're not isolating the filesystem, so I don't see that
> there's much to gain by isolating pids.  I'd be happy enough to see
> CLONE_NEWPID removed unconditionally.

The points are described in the message for 0515adceaa8f
("passt, pasta: Namespace-based sandboxing, defer seccomp policy
application"):

- no need to track child processes: when PID 1 exits, they all
  terminate, instead of that crazy / buggy hack I implemented before
  this change

- for passt, to avoid possible attacks based on the knowledge of a
  target process' PID, but then I realised that would also apply to
  pasta's spawned process

Isolating the filesystem would make it almost entirely useless for all
the nice debugging / development features it offers (like running any
network utility or test), but there's not much we would gain in a
general case by keeping the original PID namespace (you don't use pasta
to run strace or gdb), so I think it's an easy, albeit small, win from
a security perspective. I'd rather keep it.

-- 
Stefano