Re: [PATCH] Enable SR-IOV instantiation through /sys file

["Duyck, Alexander H" <alexander.h.duyck@intel.com>]

On Sat, 2017-10-28 at 00:19 +0200, Alex Williamson wrote:
> On Fri, 27 Oct 2017 21:50:43 +0000
> "Wang, Liang-min" <liang-min.wang@intel.com> wrote:
> 
> > > -----Original Message-----
> > > From: Alex Williamson [mailto:alex.williamson@redhat.com]
> > > Sent: Tuesday, October 24, 2017 6:07 PM
> > > To: Wang, Liang-min <liang-min.wang@intel.com>
> > > Cc: Kirsher, Jeffrey T <jeffrey.t.kirsher@intel.com>; kvm@vger.kernel.org;
> > > linux-pci@vger.kernel.org; linux-kernel@vger.kernel.org;
> > > bhelgaas@google.com; Duyck, Alexander H <alexander.h.duyck@intel.com>
> > > Subject: Re: [PATCH] Enable SR-IOV instantiation through /sys file
> > > 
> > > On Tue, 24 Oct 2017 21:49:15 +0000
> > > "Wang, Liang-min" <liang-min.wang@intel.com> wrote:
> > >   
> > > > Just like any PCIe devices that supports SR-IOV. There are restrictions set for  
> > > 
> > > VF. Also, there is a concept of trust VF now available for PF to manage certain
> > > features that only selected VF could exercise. Are you saying all the devices
> > > supporting SR-IOV all have security issue?
> > > 
> > > Here's a simple example, most SR-IOV capable NICs, including those from
> > > Intel, require the PF interface to be up in order to route traffic from
> > > the VF.  If the user controls the PF interface and VFs are used
> > > elsewhere in the host, the PF driver in userspace can induce a denial
> > > of service on the VFs.  That doesn't even take into account that VFs
> > > might be in separate IOMMU groups from the PF and therefore not
> > > isolated from the host like the PF and that the PF driver can
> > > potentially manipulate the VF, possibly performing DMA on behalf of the
> > > PF.  VFs are only considered secure today because the PF is managed by
> > > a driver in the host kernel.  Allowing simple enablement of VFs for a
> > > user owned PF seems inherently insecure to me.  Thanks,
> > > 
> > > Alex  
> > 
> > Firstly, the concern is on user-space PF driver based upon vfio-pci, this patch doesn't
> > change PF behavior so with/without this patch, the concern remains the same.
> 
> This patch enables SR-IOV to be enabled via the host on a user-owned
> PF, how is this not a change in behavior?
> 
> > Secondly, the security concern (including denial of service) in general is to ensure trust
> > entity to be trust-worthy. No matter the PF driver is in kernel-space or in user- space,
> > necessary mechanism needs to be enforced on the device driver to ensure it's
> > trusted worthy. For example, ixgbe kernel driver introduces a Tx hang detection
> > to avoid driver stays in a bad state. Therefore, it's the responsibility of user-space
> > driver function, which based upon vfio-pci, to enforce necessary mechanism to ensure
> > its trust-ness. That's a given.
> 
> Userspace is not trustworthy, therefore the host kernel cannot place
> responsibility on a userspace driver for anything, including the
> behavior of VFs.  I'm sorry, but it's a NAK unless you intend to
> follow-up with some proposal to quarantine the VFs enabled by the
> userspace PF driver.  Thanks,
> 
> Alex

I don't see this so much as a security problem per-se. It all depends
on the hardware setup. If I recall correctly, there are devices where
the PF function doesn't really do much other than act as a bit more
heavy-weight VF, and the actual logic is handled by a firmware engine
on the device. The only real issue is that for devices like the Intel
NICs instead of trusting a firmware engine we have historically used a
kernel driver and now we are wanting to trust a user-space agent
instead.

I do think that we probably need to have some sort of signaling between
user-space and vfio-pci that would allow for notifying the user-space
of the change and for user-space to notify vfio-pci that it is capable
of handling the notification. This is something that can be toggled at
any time after all and not all devices have a means of notifying the PF
that this has been changed.

Beyond that once the root user enables the VFs I would kind of think
they know what driver they have running them. Enabling VFs implies the
root user trusts the application running on top of vfio-pci to handle
the PF responsibly. At least that is how it works in my mind.

Thanks.

- Alexander
  (using full name since 2 Alexs in one thread can be confusing)