Date:   Wed, 5 Feb 2020 07:13:15 -0700
From:   Alex Williamson <alex.williamson@...hat.com>
To:     "Liu, Yi L" <yi.l.liu@...el.com>
Cc:     "kvm@...r.kernel.org" <kvm@...r.kernel.org>,
        "linux-pci@...r.kernel.org" <linux-pci@...r.kernel.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        "dev@...k.org" <dev@...k.org>,
        "mtosatti@...hat.com" <mtosatti@...hat.com>,
        "thomas@...jalon.net" <thomas@...jalon.net>,
        "bluca@...ian.org" <bluca@...ian.org>,
        "jerinjacobk@...il.com" <jerinjacobk@...il.com>,
        "Richardson, Bruce" <bruce.richardson@...el.com>,
        "cohuck@...hat.com" <cohuck@...hat.com>
Subject: Re: [RFC PATCH 3/7] vfio/pci: Introduce VF token

On Wed, 5 Feb 2020 07:57:29 +0000
"Liu, Yi L" <yi.l.liu@...el.com> wrote:

> > From: Alex Williamson <alex.williamson@...hat.com>
> > Sent: Wednesday, February 5, 2020 7:06 AM
> > To: kvm@...r.kernel.org
> > Subject: [RFC PATCH 3/7] vfio/pci: Introduce VF token
> > 
> > If we enable SR-IOV on a vfio-pci owned PF, the resulting VFs are not
> > fully isolated from the PF.  The PF can always cause a denial of
> > service to the VF, if not access data passed through the VF directly.
> > This is why vfio-pci currently does not bind to PFs with SR-IOV enabled
> > and does not provide access itself to enabling SR-IOV on a PF.  The
> > IOMMU grouping mechanism might allow us a solution to this lack of
> > isolation, however the deficiency isn't actually in the DMA path, so
> > much as the potential cooperation between PF and VF devices.  Also,
> > if we were to force VFs into the same IOMMU group as the PF, we severely
> > limit the utility of having independent drivers managing PFs and VFs
> > with vfio.
> > 
> > Therefore we introduce the concept of a VF token.  The token is
> > implemented as a UUID and represents a shared secret which must be set
> > by the PF driver and used by the VF drivers in order to access a vfio
> > device file descriptor for the VF.  The ioctl to set the VF token will
> > be provided in a later commit, this commit implements the underlying
> > infrastructure.  The concept here is to augment the string the user
> > passes to match a device within a group in order to retrieve access to
> > the device descriptor.  For example, rather than passing only the PCI
> > device name (ex. "0000:03:00.0") the user would also pass a vf_token
> > UUID (ex. "2ab74924-c335-45f4-9b16-8569e5b08258").  The device match
> > string therefore becomes:
> > 
> > "0000:03:00.0 vf_token=2ab74924-c335-45f4-9b16-8569e5b08258"
> > 
> > This syntax is expected to be extensible to future options as well, with
> > the standard being:
> > 
> > "$DEVICE_NAME $OPTION1=$VALUE1 $OPTION2=$VALUE2"
> > 
> > The device name must be first and option=value pairs are separated by
> > spaces.
> > 
> > The vf_token option is only required for VFs where the PF device is
> > bound to vfio-pci.  There is no change for PFs using existing host
> > drivers.
> > 
> > Note that in order to protect existing VF users, not only is it required
> > to set a vf_token on the PF before VFs devices can be accessed, but also
> > if there are existing VF users, (re)opening the PF device must also
> > provide the current vf_token as authentication.  This is intended to
> > prevent a VF driver starting with a trusted PF driver and later being
> > replaced by an unknown driver.  A vf_token is not required to open the
> > PF device when none of the VF devices are in use by vfio-pci drivers.  
> 
> So vfio_token is a kind of per-PF token?

Yes, the token is per-PF.  Note that the token can be changed and it
does not "de-authenticate" opened VFs.  Thanks,

Alex

Powered by blists - more mailing lists