Re: [PATCH v1 2/2] vfio/pci: Emulate PASID/PRI capability for VFs

[Alex Williamson]

On Tue, 14 Apr 2020 02:40:58 +0000
"Tian, Kevin" <kevin.tian@xxxxxxxxx> wrote:

> > From: Alex Williamson <alex.williamson@xxxxxxxxxx>
> > Sent: Tuesday, April 14, 2020 3:21 AM
> > 
> > On Mon, 13 Apr 2020 08:05:33 +0000
> > "Tian, Kevin" <kevin.tian@xxxxxxxxx> wrote:
> >   
> > > > From: Tian, Kevin
> > > > Sent: Monday, April 13, 2020 3:55 PM
> > > >  
> > > > > From: Raj, Ashok <ashok.raj@xxxxxxxxxxxxxxx>
> > > > > Sent: Monday, April 13, 2020 11:11 AM
> > > > >
> > > > > On Wed, Apr 08, 2020 at 10:19:40AM -0600, Alex Williamson wrote:  
> > > > > > On Tue, 7 Apr 2020 21:00:21 -0700
> > > > > > "Raj, Ashok" <ashok.raj@xxxxxxxxx> wrote:
> > > > > >  
> > > > > > > Hi Alex
> > > > > > >
> > > > > > > + Bjorn  
> > > > > >
> > > > > >  + Don
> > > > > >  
> > > > > > > FWIW I can't understand why PCI SIG went different ways with ATS,
> > > > > > > where its enumerated on PF and VF. But for PASID and PRI its only
> > > > > > > in PF.
> > > > > > >
> > > > > > > I'm checking with our internal SIG reps to followup on that.
> > > > > > >
> > > > > > > On Tue, Apr 07, 2020 at 09:58:01AM -0600, Alex Williamson wrote:  
> > > > > > > > > Is there vendor guarantee that hidden registers will locate at the
> > > > > > > > > same offset between PF and VF config space?  
> > > > > > > >
> > > > > > > > I'm not sure if the spec really precludes hidden registers, but the
> > > > > > > > fact that these registers are explicitly outside of the capability
> > > > > > > > chain implies they're only intended for device specific use, so I'd  
> > say  
> > > > > > > > there are no guarantees about anything related to these registers.  
> > > > > > >
> > > > > > > As you had suggested in the other thread, we could consider
> > > > > > > using the same offset as in PF, but even that's a better guess
> > > > > > > still not reliable.
> > > > > > >
> > > > > > > The other option is to maybe extend driver ops in the PF to expose
> > > > > > > where the offsets should be. Sort of adding the quirk in the
> > > > > > > implementation.
> > > > > > >
> > > > > > > I'm not sure how prevalent are PASID and PRI in VF devices. If SIG is  
> > > > > resisting  
> > > > > > > making VF's first class citizen, we might ask them to add some  
> > verbiage  
> > > > > > > to suggest leave the same offsets as PF open to help emulation  
> > software.  
> > > > > >
> > > > > > Even if we know where to expose these capabilities on the VF, it's not
> > > > > > clear to me how we can actually virtualize the capability itself.  If
> > > > > > the spec defines, for example, an enable bit as r/w then software that
> > > > > > interacts with that register expects the bit is settable.  There's no
> > > > > > protocol for "try to set the bit and re-read it to see if the hardware
> > > > > > accepted it".  Therefore a capability with a fixed enable bit
> > > > > > representing the state of the PF, not settable by the VF, is
> > > > > > disingenuous to the spec.  
> > > > >
> > > > > I think we are all in violent agreement. A lot of times the pci spec gets
> > > > > defined several years ahead of real products and no one remembers
> > > > > the justification on why they restricted things the way they did.
> > > > >
> > > > > Maybe someone early product wasn't quite exposing these features to  
> > the  
> > > > > VF
> > > > > and hence the spec is bug compatible :-)
> > > > >  
> > > > > >
> > > > > > If what we're trying to do is expose that PASID and PRI are enabled on
> > > > > > the PF to a VF driver, maybe duplicating the PF capabilities on the VF
> > > > > > without the ability to control it is not the right approach.  Maybe we  
> > > > >
> > > > > As long as the capability enable is only provided when the PF has  
> > enabled  
> > > > > the feature. Then it seems the hardware seems to do the right thing.
> > > > >
> > > > > Assume we expose PASID/PRI only when PF has enabled it. It will be the
> > > > > case since the PF driver needs to exist, and IOMMU would have set the
> > > > > PASID/PRI/ATS on PF.
> > > > >
> > > > > If the emulation is purely spoofing the capability. Once vIOMMU driver
> > > > > enables PASID, the context entries for the VF are completely  
> > independent  
> > > > > from the PF context entries.
> > > > >
> > > > > vIOMMU would enable PASID, and we just spoof the PASID capability.
> > > > >
> > > > > If vIOMMU or guest for some reason does disable_pasid(), then the
> > > > > vIOMMU driver can disaable PASID on the VF context entries. So the VF
> > > > > although the capability is blanket enabled on PF, IOMMU gaurantees  
> > the  
> > > > > transactions are blocked.
> > > > >
> > > > >
> > > > > In the interim, it seems like the intent of the virtual capability
> > > > > can be honored via help from the IOMMU for the controlling aspect..
> > > > >
> > > > > Did i miss anything?  
> > > >
> > > > Above works for emulating the enable bit (under the assumption that
> > > > PF driver won't disable pasid when vf is assigned). However, there are
> > > > also "Execute permission enable" and "Privileged mode enable" bits in
> > > > PASID control registers. I don't know how those bits could be cleanly
> > > > emulated when the guest writes a value different from PF's...  
> > >
> > > sent too quick. the IOMMU also includes control bits for allowing/
> > > blocking execute requests and supervisor requests. We can rely on
> > > IOMMU to block those requests to emulate the disabled cases of
> > > all three control bits in the pasid cap.  
> > 
> > 
> > So if the emulation of the PASID capability takes into account the
> > IOMMU configuration to back that emulation, shouldn't we do that
> > emulation in the hypervisor, ie. QEMU, rather than the kernel vfio
> > layer?  Thanks,
> > 
> > Alex  
> 
> We need enforce it in physical IOMMU, to ensure that even the
> VF may send requests which violate the guest expectation those
> requests are always blocked by IOMMU. Kernel vfio identifies
> such need when emulating the pasid cap and then forward the 
> request to host iommu driver.

Implementing this in the kernel would be necessary if we needed to
protect from the guest device doing something bad to the host or
other devices.  Making sure the physical IOMMU is configured to meet
guest expectations doesn't sound like it necessarily falls into that
category.  We do that on a regular basis to program the DMA mappings.
Tell me more about why the hypervisor can't handle this piece of
guest/host synchronization on top of all the other things it
synchronizes to make a VM.  Thanks,

Alex