Re: [PATCH] iommu: Relax ACS requirement for RCiEP devices.

[Raj, Ashok]

Hi Alex,

I was able to find better language in the IOMMU spec that gaurantees 
the behavior we need. See below.


On Tue, May 05, 2020 at 09:34:14AM -0600, Alex Williamson wrote:
> On Tue, 5 May 2020 07:56:06 -0700
> "Raj, Ashok" <ashok.raj@xxxxxxxxx> wrote:
> 
> > On Tue, May 05, 2020 at 08:05:14AM -0600, Alex Williamson wrote:
> > > On Mon, 4 May 2020 23:11:07 -0700
> > > "Raj, Ashok" <ashok.raj@xxxxxxxxx> wrote:
> > >   
> > > > Hi Alex
> > > > 
> > > > + Joerg, accidently missed in the Cc.
> > > > 
> > > > On Mon, May 04, 2020 at 11:19:36PM -0600, Alex Williamson wrote:  
> > > > > On Mon,  4 May 2020 21:42:16 -0700
> > > > > Ashok Raj <ashok.raj@xxxxxxxxx> wrote:
> > > > >     
> > > > > > PCIe Spec recommends we can relax ACS requirement for RCIEP devices.
> > > > > > 
> > > > > > PCIe 5.0 Specification.
> > > > > > 6.12 Access Control Services (ACS)
> > > > > > Implementation of ACS in RCiEPs is permitted but not required. It is
> > > > > > explicitly permitted that, within a single Root Complex, some RCiEPs
> > > > > > implement ACS and some do not. It is strongly recommended that Root Complex
> > > > > > implementations ensure that all accesses originating from RCiEPs
> > > > > > (PFs and VFs) without ACS capability are first subjected to processing by
> > > > > > the Translation Agent (TA) in the Root Complex before further decoding and
> > > > > > processing. The details of such Root Complex handling are outside the scope
> > > > > > of this specification.
> > > > > >   
> > > > > 
> > > > > Is the language here really strong enough to make this change?  ACS is
> > > > > an optional feature, so being permitted but not required is rather
> > > > > meaningless.  The spec is also specifically avoiding the words "must"
> > > > > or "shall" and even when emphasized with "strongly", we still only have
> > > > > a recommendation that may or may not be honored.  This seems like a
> > > > > weak basis for assuming that RCiEPs universally honor this
> > > > > recommendation.  Thanks,
> > > > >     
> > > > 
> > > > We are speaking about PCIe spec, where people write it about 5 years ahead
> > > > and every vendor tries to massage their product behavior with vague
> > > > words like this..  :)
> > > > 
> > > > But honestly for any any RCiEP, or even integrated endpoints, there 
> > > > is no way to send them except up north. These aren't behind a RP.  
> > > 
> > > But they are multi-function devices and the spec doesn't define routing
> > > within multifunction packages.  A single function RCiEP will already be
> > > assumed isolated within its own group.  
> > 
> > That's right. The other two devices only have legacy PCI headers. So 
> > they can't claim to be RCiEP's but just integrated endpoints. The legacy
> > devices don't even have a PCIe header.
> > 
> > I honestly don't know why these are groped as MFD's in the first place.
> > 
> > >    
> > > > I did check with couple folks who are part of the SIG, and seem to agree
> > > > that ACS treatment for RCiEP's doesn't mean much. 
> > > > 
> > > > I understand the language isn't strong, but it doesn't seem like ACS should
> > > > be a strong requirement for RCiEP's and reasonable to relax.
> > > > 
> > > > What are your thoughts?   
> > > 
> > > I think hardware vendors have ACS at their disposal to clarify when
> > > isolation is provided, otherwise vendors can submit quirks, but I don't
> > > see that the "strongly recommended" phrasing is sufficient to assume
> > > isolation between multifunction RCiEPs.  Thanks,  
> > 
> > You point is that integrated MFD endpoints, without ACS, there is no 
> > gaurantee to SW that they are isolated.
> > 
> > As far as a quirk, do you think:
> > 	- a cmdline optput for integrated endpoints, and RCiEP's suffice?
> > 	  along with a compile time default that is strict enforcement
> > 	- typical vid/did type exception list?
> > 
> > A more generic way to ask for exception would be scalable until we can stop
> > those type of integrated devices. Or we need to maintain these device lists
> > for eternity. 
> 
> I don't think the language in the spec is anything sufficient to handle
> RCiEP uniquely.  We've previously rejected kernel command line opt-outs
> for ACS, and the extent to which those patches still float around the
> user community and are blindly used to separate IOMMU groups are a
> testament to the failure of this approach.  Users do not have a basis
> for enabling this sort of opt-out.  The benefit is obvious in the IOMMU
> grouping, but the risk is entirely unknown.  A kconfig option is even
> worse as that means if you consume a downstream kernel, the downstream
> maintainers might have decided universally that isolation is less
> important than functionality.

We discussed this internally, and Intel vt-d spec does spell out clearly 
in Section 3.16 Root-Complex Peer to Peer Considerations. The spec clearly
calls out that all p2p must be done on translated addresses and therefore
must go through the IOMMU.

I suppose they should also have some similar platform gauranteed behavior
for RCiEP's or MFD's *Must* behave as follows. The language is strict and
when IOMMU is enabled in the platform, everything is sent up north to the
IOMMU agent.

3.16 Root-Complex Peer to Peer Considerations
When DMA remapping is enabled, peer-to-peer requests through the
Root-Complex must be handled
as follows:
â The input address in the request is translated (through first-level,
  second-level or nested translation) to a host physical address (HPA).
  The address decoding for peer addresses must be done only on the 
  translated HPA. Hardware implementations are free to further limit 
  peer-to-peer accesses to specific host physical address regions 
  (or to completely disallow peer-forwarding of translated requests).
â Since address translation changes the contents (address field) of the PCI
  Express Transaction Layer Packet (TLP), for PCI Express peer-to-peer 
  requests with ECRC, the Root-Complex hardware must use the new ECRC 
  (re-computed with the translated address) if it decides to forward 
  the TLP as a peer request.
â Root-ports, and multi-function root-complex integrated endpoints, may
  support additional peerto-peer control features by supporting PCI Express
  Access Control Services (ACS) capability. Refer to ACS capability in 
  PCI Express specifications for details.

> to indicate where devices are isolated.  The hardware can do this
> itself by implementing ACS, otherwise we need quirks.  I think we've
> also generally been reluctant to accept quirks that provide a blanket
> opt-out for a vendor because doing so is akin to trying to predict the
> future (determining the behavior of all current and previous hardware
> is generally a sufficiently impossible task already).  Perhaps if a
> vendor has a published internal policy regarding RCiEP isolation and is
> willing to stand by a quirk, there might be room to negotiate.  Thanks,
> 
> Alex
>