[PATCH 5/5] vfio-pci: Allow to mmap MSI-X table if interrupt remapping is supported

[Tian, Kevin]

> From: Alex Williamson [mailto:alex.williamson at redhat.com]
> Sent: Thursday, May 12, 2016 10:21 AM
> 
> On Thu, 12 May 2016 01:19:44 +0000
> "Tian, Kevin" <kevin.tian at intel.com> wrote:
> 
> > > From: Alex Williamson [mailto:alex.williamson at redhat.com]
> > > Sent: Wednesday, May 11, 2016 11:54 PM
> > >
> > > On Wed, 11 May 2016 06:29:06 +0000
> > > "Tian, Kevin" <kevin.tian at intel.com> wrote:
> > >
> > > > > From: Alex Williamson [mailto:alex.williamson at redhat.com]
> > > > > Sent: Thursday, May 05, 2016 11:05 PM
> > > > >
> > > > > On Thu, 5 May 2016 12:15:46 +0000
> > > > > "Tian, Kevin" <kevin.tian at intel.com> wrote:
> > > > >
> > > > > > > From: Yongji Xie [mailto:xyjxie at linux.vnet.ibm.com]
> > > > > > > Sent: Thursday, May 05, 2016 7:43 PM
> > > > > > >
> > > > > > > Hi David and Kevin,
> > > > > > >
> > > > > > > On 2016/5/5 17:54, David Laight wrote:
> > > > > > >
> > > > > > > > From: Tian, Kevin
> > > > > > > >> Sent: 05 May 2016 10:37
> > > > > > > > ...
> > > > > > > >>> Acutually, we are not aimed at accessing MSI-X table from
> > > > > > > >>> guest. So I think it's safe to passthrough MSI-X table if we
> > > > > > > >>> can make sure guest kernel would not touch MSI-X table in
> > > > > > > >>> normal code path such as para-virtualized guest kernel on PPC64.
> > > > > > > >>>
> > > > > > > >> Then how do you prevent malicious guest kernel accessing it?
> > > > > > > > Or a malicious guest driver for an ethernet card setting up
> > > > > > > > the receive buffer ring to contain a single word entry that
> > > > > > > > contains the address associated with an MSI-X interrupt and
> > > > > > > > then using a loopback mode to cause a specific packet be
> > > > > > > > received that writes the required word through that address.
> > > > > > > >
> > > > > > > > Remember the PCIe cycle for an interrupt is a normal memory write
> > > > > > > > cycle.
> > > > > > > >
> > > > > > > > 	David
> > > > > > > >
> > > > > > >
> > > > > > > If we have enough permission to load a malicious driver or
> > > > > > > kernel, we can easily break the guest without exposed
> > > > > > > MSI-X table.
> > > > > > >
> > > > > > > I think it should be safe to expose MSI-X table if we can
> > > > > > > make sure that malicious guest driver/kernel can't use
> > > > > > > the MSI-X table to break other guest or host. The
> > > > > > > capability of IRQ remapping could provide this
> > > > > > > kind of protection.
> > > > > > >
> > > > > >
> > > > > > With IRQ remapping it doesn't mean you can pass through MSI-X
> > > > > > structure to guest. I know actual IRQ remapping might be platform
> > > > > > specific, but at least for Intel VT-d specification, MSI-X entry must
> > > > > > be configured with a remappable format by host kernel which
> > > > > > contains an index into IRQ remapping table. The index will find a
> > > > > > IRQ remapping entry which controls interrupt routing for a specific
> > > > > > device. If you allow a malicious program random index into MSI-X
> > > > > > entry of assigned device, the hole is obvious...
> > > > > >
> > > > > > Above might make sense only for a IRQ remapping implementation
> > > > > > which doesn't rely on extended MSI-X format (e.g. simply based on
> > > > > > BDF). If that's the case for PPC, then you should build MSI-X
> > > > > > passthrough based on this fact instead of general IRQ remapping
> > > > > > enabled or not.
> > > > >
> > > > > I don't think anyone is expecting that we can expose the MSI-X vector
> > > > > table to the guest and the guest can make direct use of it.  The end
> > > > > goal here is that the guest on a power system is already
> > > > > paravirtualized to not program the device MSI-X by directly writing to
> > > > > the MSI-X vector table.  They have hypercalls for this since they
> > > > > always run virtualized.  Therefore a) they never intend to touch the
> > > > > MSI-X vector table and b) they have sufficient isolation that a guest
> > > > > can only hurt itself by doing so.
> > > > >
> > > > > On x86 we don't have a), our method of programming the MSI-X vector
> > > > > table is to directly write to it. Therefore we will always require QEMU
> > > > > to place a MemoryRegion over the vector table to intercept those
> > > > > accesses.  However with interrupt remapping, we do have b) on x86, which
> > > > > means that we don't need to be so strict in disallowing user accesses
> > > > > to the MSI-X vector table.  It's not useful for configuring MSI-X on
> > > > > the device, but the user should only be able to hurt themselves by
> > > > > writing it directly.  x86 doesn't really get anything out of this
> > > > > change, but it helps this special case on power pretty significantly
> > > > > aiui.  Thanks,
> > > > >
> > > >
> > > > Allowing guest direct write to MSI-x table has system-wide impact.
> > > > As I explained earlier, hypervisor needs to control "interrupt_index"
> > > > programmed in MSI-X entry, which is used to associate a specific
> > > > IRQ remapping entry. Now if you expose whole MSI-x table to guest,
> > > > it can program random index into MSI-X entry to associate with
> > > > any IRQ remapping entry and then there won't be any isolation per se.
> > > >
> > > > You can check "5.5.2 MSI and MSI-X Register Programming" in VT-d
> > > > spec.
> > >
> > > I think you're extrapolating beyond the vfio interface.  The change
> > > here is to remove the vfio protection of the MSI-X vector table when
> > > the system provides interrupt isolation.  The argument is that this is
> > > safe to do because the hardware protects the host from erroneous and
> > > malicious user programming, but it is not meant to provide a means to
> > > program MSI-X directly through the vector table.  This is effectively
> >
> > Sorry I didn't get this point. Once we allow userspace to mmap MSI-X
> > table, isn't it equivalent to allowing userspace directly program vector
> > table? Or is there other mechanism to prevent direct programming?
> 
> This would remove the mechanism that prevents direct programming.
> Users can configure the device to perform DMA w/o setting up an IOMMU
> mapping, which generates a fault.  Users can incorrectly manipulate the
> MSI-X vector table instead of using the proper vfio programming
> interface, which generates a fault.
> 
> > > the same as general DMA programming, if the vfio programming model is
> > > not followed the device generates iommu faults.  I do have a concern
> > > that userspace driver writers are going to more often presume they can
> > > use the vector table directly because of this change, but I don't know
> > > that that is sufficient reason to prevent such a change.  They'll
> > > quickly discover the device generates faults on interrupt rather than
> > > working as expected.
> >
> > If userspace can actually program vector table directly, there is not
> > always fault triggered. As long as MSI-X table is fully under control
> > of userspace, any interrupt index can be used here which may link
> > to a IRQ remapping entry allocated for other devices.
> 
> Is not part of the vector lookup comparing the source ID of the DMA
> write?  If not then VT-d interrupt remapping offers us no protection
> from a malicious guest performing DMA to the same address.  Thanks,
> 

For x86 IRQ remapping doesn't rely on existing DMAR remapping table for
DMA. Instead IRQ remapping table is a global table per VT-d engine, shared
by all devices behind this VT-d engine. Each IRQ remapping table entry 
(IRTE) can specify:

- whether to validate source-id of the interrupt requests;
- whether to verify the whole source id;
- whether to verify some bits of the source id;
...
(section 9.10 Interrupt Remapping Table Entry (IRTE) in VT-d spec)

IRTE index is programmed into corresponding MSI-X entry of a device.
When an interrupt message is triggered from that device, IRQ remapping
hardware will use the index to find the IRTE entry.

As long as host kernel has strict control of MSI-X table, it can choose to
not validate source-id as long as IRTE index is trusted.

That is why I concerned you cannot do this simply based on whether 
IRQ remapping is available on x86. If such usage is really required,
you'll need some more accurate indicator per iommu structure to tell
whether safe to allow mmap MSI-X to user space.

Thanks
Kevin