Re: [RFC 0/8] PGP key parser using SandBox Mode

[Roberto Sassu]

On 2/16/2024 6:21 PM, Jonathan Corbet wrote:
> Petr Tesařík <petr@xxxxxxxxxxx> writes:
>
> > On Fri, 16 Feb 2024 07:38:30 -0800
> > Dave Hansen <dave.hansen@xxxxxxxxx> wrote:
> > > I'm confused by this.  The kernel doesn't (appear to) have a PGP parser
> > > today.  So are you saying that it *should* have one and it's only
> > > feasible if its confined in a sandbox?
> >
> > I'm sorry if this is confusing. Yes, your understanding is correct.
> > This patch series demonstrates that SBM (even in the initial version
> > that was submitted) allows to write a PGP parser which can survive
> > memory safety bugs withoug compromising the rest of the kernel.
>
> So I have a different question: some years ago we added the "usermode
> blob" feature for just this kind of use case - parsing firewall rules at
> the time.  It has never been used for that, but it's still there in
> kernel/usermode_driver.c.  Is there a reason why this existing
> functionality can't be used for tasks like PGP parsing as well?

Yes, it was an option I explored last year (briefly talked about it as a 
BoF at LSS NA 2023).

You are right, there is such feature that seemed to fit well.

User space blob embedded in a kernel module, so signed. User space 
process connected only to the kernel through a pipe.

I even went ahead, and created a framework:

https://lore.kernel.org/linux-kernel/20230317145240.363908-1-roberto.sassu@xxxxxxxxxxxxxxx/

so that anyone can implement similar use cases.

The further step is: how can I ensure that the process launched by the 
kernel is not attacked by root (which I assumed to be less powerful than 
the kernel in a locked-down system).

I handled this in both directions:

- The process launched by the kernel is under a seccomp strict profile,
  and can only read/write through the pipe created by the kernel (and
  call few other system calls, such as exit()). Otherwise it is killed.
  Cannot create any new communication channel.

- I created an LSM that denies any attempt to ptrace/signal to the
  process launched by the kernel. Jann Horn also suggested to make the
  process non-swappable.

However, despite these attempts, security people don't feel confident on 
offloading a kernel workload outside the kernel.

This is why this work started.

Roberto