RE: [PATCH v2 0/7] vfio/pci: SR-IOV support

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



> From: Alex Williamson
> Sent: Thursday, February 20, 2020 2:54 AM
> 
> Changes since v1 are primarily to patch 3/7 where the commit log is
> rewritten, along with option parsing and failure logging based on
> upstream discussions.  The primary user visible difference is that
> option parsing is now much more strict.  If a vf_token option is
> provided that cannot be used, we generate an error.  As a result of
> this, opening a PF with a vf_token option will serve as a mechanism of
> setting the vf_token.  This seems like a more user friendly API than
> the alternative of sometimes requiring the option (VFs in use) and
> sometimes rejecting it, and upholds our desire that the option is
> always either used or rejected.
> 
> This also means that the VFIO_DEVICE_FEATURE ioctl is not the only
> means of setting the VF token, which might call into question whether
> we absolutely need this new ioctl.  Currently I'm keeping it because I
> can imagine use cases, for example if a hypervisor were to support
> SR-IOV, the PF device might be opened without consideration for a VF
> token and we'd require the hypservisor to close and re-open the PF in
> order to set a known VF token, which is impractical.
> 
> Series overview (same as provided with v1):

Thanks for doing this! 

> 
> The synopsis of this series is that we have an ongoing desire to drive
> PCIe SR-IOV PFs from userspace with VFIO.  There's an immediate need
> for this with DPDK drivers and potentially interesting future use

Can you provide a link to the DPDK discussion? 

> cases in virtualization.  We've been reluctant to add this support
> previously due to the dependency and trust relationship between the
> VF device and PF driver.  Minimally the PF driver can induce a denial
> of service to the VF, but depending on the specific implementation,
> the PF driver might also be responsible for moving data between VFs
> or have direct access to the state of the VF, including data or state
> otherwise private to the VF or VF driver.

Just a loud thinking. While the motivation of VF token sounds reasonable
to me, I'm curious why the same concern is not raised in other usages.
For example, there is no such design in virtio framework, where the
virtio device could also be restarted, putting in separate process (vhost-user),
and even in separate VM (virtio-vhost-user), etc. Of course the para-
virtualized attribute of virtio implies some degree of trust, but as you
mentioned many SR-IOV implementations support VF->PF communication
which also implies some level of trust. It's perfectly fine if VFIO just tries
to do better than other sub-systems, but knowing how other people
tackle the similar problem may make the whole picture clearer. 😊

+Jason.

> 
> To help resolve these concerns, we introduce a VF token into the VFIO
> PCI ABI, which acts as a shared secret key between drivers.  The
> userspace PF driver is required to set the VF token to a known value
> and userspace VF drivers are required to provide the token to access
> the VF device.  If a PF driver is restarted with VF drivers in use, it
> must also provide the current token in order to prevent a rogue
> untrusted PF driver from replacing a known driver.  The degree to
> which this new token is considered secret is left to the userspace
> drivers, the kernel intentionally provides no means to retrieve the
> current token.

I'm wondering whether the token idea can be used beyond SR-IOV, e.g.
(1) we may allow vfio user space to manage Scalable IOV in the future,
which faces the similar challenge between the PF and mdev; (2) the
token might be used as a canonical way to replace off-tree acs-override
workaround, say, allowing the admin to assign devices within the 
same iommu group to different VMs which trust each other. I'm not
sure how much complexity will be further introduced, but it's greatly
appreciated if you can help think a bit and if feasible abstract some 
logic in vfio core layer for such potential usages...

> 
> Note that the above token is only required for this new model where
> both the PF and VF devices are usable through vfio-pci.  Existing
> models of VFIO drivers where the PF is used without SR-IOV enabled
> or the VF is bound to a userspace driver with an in-kernel, host PF
> driver are unaffected.
> 
> The latter configuration above also highlights a new inverted scenario
> that is now possible, a userspace PF driver with in-kernel VF drivers.
> I believe this is a scenario that should be allowed, but should not be
> enabled by default.  This series includes code to set a default
> driver_override for VFs sourced from a vfio-pci user owned PF, such
> that the VFs are also bound to vfio-pci.  This model is compatible
> with tools like driverctl and allows the system administrator to
> decide if other bindings should be enabled.  The VF token interface
> above exists only between vfio-pci PF and VF drivers, once a VF is
> bound to another driver, the administrator has effectively pronounced
> the device as trusted.  The vfio-pci driver will note alternate
> binding in dmesg for logging and debugging purposes.
> 
> Please review, comment, and test.  The example QEMU implementation
> provided with the RFC is still current for this version.  Thanks,
> 
> Alex
> 
> RFC:
> https://lore.kernel.org/lkml/158085337582.9445.17682266437583505502.stg
> it@xxxxxxxxxx/
> v1:
> https://lore.kernel.org/lkml/158145472604.16827.15751375540102298130.st
> git@xxxxxxxxxx/
> 
> ---
> 
> Alex Williamson (7):
>       vfio: Include optional device match in vfio_device_ops callbacks
>       vfio/pci: Implement match ops
>       vfio/pci: Introduce VF token
>       vfio: Introduce VFIO_DEVICE_FEATURE ioctl and first user
>       vfio/pci: Add sriov_configure support
>       vfio/pci: Remove dev_fmt definition
>       vfio/pci: Cleanup .probe() exit paths
> 
> 
>  drivers/vfio/pci/vfio_pci.c         |  383
> +++++++++++++++++++++++++++++++++--
>  drivers/vfio/pci/vfio_pci_private.h |   10 +
>  drivers/vfio/vfio.c                 |   20 +-
>  include/linux/vfio.h                |    4
>  include/uapi/linux/vfio.h           |   37 +++
>  5 files changed, 426 insertions(+), 28 deletions(-)





[Index of Archives]     [KVM ARM]     [KVM ia64]     [KVM ppc]     [Virtualization Tools]     [Spice Development]     [Libvirt]     [Libvirt Users]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite Questions]     [Linux Kernel]     [Linux SCSI]     [XFree86]

  Powered by Linux