Re: [virtio-dev] [PATCH] virtio-iommu: Rework the bypass feature

[Jean-Philippe Brucker <jean-philippe@linaro.org>]

Hi Eric,

On Wed, Oct 06, 2021 at 02:53:50PM +0200, Eric Auger wrote:
> > * The hypervisor can start the VM with bypass enabled or, if it knows
> >   that the software stack supports it, disabled. The 'bypass' config
> >   fields resets to 0 or 1.
> I don't see the reset value documented in the actual spec while it looks
> quite important for our boot bypass functionality. Or did I miss it?

No you're right, I'll add this. It's important to say that the device may
initialize the field to 1, but also that a reset (writing a 0 to device
status) doesn't re-initialize the field to its boot value (because that
may introduce a vulnerability during the FW->OS transition)

  If the device offers the VIRTIO_IOMMU_F_BYPASS_CONFIG feature, it MAY
  initialize the \field{bypass} field to 1. When the driver resets the
  device, the \field{bypass} field SHOULD NOT change.

[...]
> > diff --git a/virtio-iommu.tex b/virtio-iommu.tex
> > index efa735b..a2c90ea 100644
> > --- a/virtio-iommu.tex
> > +++ b/virtio-iommu.tex
> > @@ -59,14 +59,19 @@ \subsection{Feature bits}\label{sec:Device Types / IOMMU Device / Feature bits}
> >    VIRTIO_IOMMU_F_MAP_UNMAP is supported.}
> >  
> >  \item[VIRTIO_IOMMU_F_BYPASS (3)]
> > -  When not attached to a domain, endpoints downstream of the IOMMU
> > -  can access the guest-physical address space.
> > +  This feature is deprecated.
> Why are you obliged to deprecate the feature.

We could leave the feature, but it adds complexity in the spec, and in
drivers if they have to support both. Maybe we can leave the feature,
merge its description with F_BYPASS_CONFIG and state something like:

  The VIRTIO_IOMMU_F_BYPASS_CONFIG feature supersedes
  VIRTIO_IOMMU_F_BYPASS. New devices SHOULD NOT offer
  VIRTIO_IOMMU_F_BYPASS.

> Can't you leave it and
> impose that either
> 
> VIRTIO_IOMMU_F_BYPASS or VIRTIO_IOMMU_F_BYPASS_CONFIG gets set?

Just to be clear, you mean devices must not implement both features at the
same time, right?  Not that devices must implement at least one of the
feature? (I don't think we can do that)

> 
> >  
> >  \item[VIRTIO_IOMMU_F_PROBE (4)]
> >    The PROBE request is available.
> >  
> >  \item[VIRTIO_IOMMU_F_MMIO (5)]
> >    The VIRTIO_IOMMU_MAP_F_MMIO flag is available.
> > +
> > +\item[VIRTIO_IOMMU_F_BYPASS_CONFIG (6)]
> > +  The global bypass state is described in \field{bypass} of
> > +  struct virtio_iommu_config. The domain bypass state is
> > +  described by VIRTIO_IOMMU_ATTACH_F_BYPASS.
> I would re-emphasize the global bypass state relates to bypass state of
> unattached devices whereas domain bypass state controls bypass for
> devices attached to a given domain.

Right, though I think I'll drop "bypass state" and describe everything
using "endpoint in bypass mode", so we can keep a single definition in one
place.

[...]
> >  \subsection{Device initialization}\label{sec:Device Types / IOMMU Device / Device initialization}
> >  
> >  When the device is reset, endpoints are not attached to any domain.
> >  
> > -If the VIRTIO_IOMMU_F_BYPASS feature is negotiated, all accesses from
> > -unattached endpoints are allowed and translated by the IOMMU using the
> > -identity function. If the feature is not negotiated, any memory access
> > -from an unattached endpoint fails. Upon attaching an endpoint in
> > -bypass mode to a new domain, any memory access from the endpoint fails,
> > -since the domain does not contain any mapping.
> > +An endpoint is in bypass mode if the VIRTIO_IOMMU_F_BYPASS_CONFIG
> > +feature is offered and:
> > +\begin{itemize}
> > +  \item config field \field{bypass} is 1 and the endpoint is not
> > +    attached to a domain, or
> > +  \item the endpoint is attached to a domain with
> > +    VIRTIO_IOMMU_ATTACH_F_BYPASS.
> > +\end{itemize}
> > +
> > +All accesses from an endpoint in bypass mode are allowed and
> > +translated by the IOMMU using the identity function.

I think I'll move this description to the Device operations section, not
sure why it was here in the first place.

> >  
> >  Future devices might support more modes of operation besides MAP/UNMAP.
> >  Drivers verify that devices set VIRTIO_IOMMU_F_MAP_UNMAP and fail
> > @@ -136,8 +156,8 @@ \subsection{Device initialization}\label{sec:Device Types / IOMMU Device / Devic
> >  
> >  \devicenormative{\subsubsection}{Device Initialization}{Device Types / IOMMU Device / Device Initialization}
> >  
> > -If the driver does not accept the VIRTIO_IOMMU_F_BYPASS feature, the
> > -device SHOULD NOT let endpoints access the guest-physical address space.
> > +The device SHOULD NOT let unattached endpoints that are not in
> > +bypass mode access the guest-physical address space.
> SHOULD NOT or MUST NOT. Can you remind me of the difference?

They are defined in rfc2119 https://datatracker.ietf.org/doc/html/rfc2119

SHOULD NOT
  This phrase, or the phrase "NOT RECOMMENDED" mean that there may exist
  valid reasons in particular circumstances when the particular behavior
  is acceptable or even useful, but the full implications should be
  understood and the case carefully weighed before implementing any
  behavior described with this label.

MUST NOT
  This phrase, or the phrase "SHALL NOT", mean that the definition is an
  absolute prohibition of the specification

So for this particular spec statement, I was reluctant to make it MUST
because there may be cases where the unattached endpoint has to access
some memory, even though I couldn't think of one. For example, maybe an
endpoint that should not perform DMA could still be allowed to send MSIs?

For the driver implementation, when the device has a "SHOULD NOT" and
doesn't follow the rule I try to handle it more gracefully, than when it
has a MUST NOT.

[...]
> >  Detach an endpoint from a domain. When this request completes, the
> > -endpoint cannot access any mapping from that domain anymore. If feature
> > -VIRTIO_IOMMU_F_BYPASS has been negotiated, then once this request
> > -completes all accesses from the endpoint are allowed and translated by the
> > -IOMMU using the identity function.
> > +endpoint cannot access any mapping from that domain anymore.
> You could adapt the rest of the paragraph to the new
> 
> VIRTIO_IOMMU_F_BYPASS_CONFIG
> 

I think the paragraph is redundant if we explain properly in "Device
operations" how bypass works.

> >  
> >  After all endpoints have been successfully detached from a domain, it
> >  ceases to exist and its ID can be reused by the driver for another domain.
> > @@ -433,6 +466,8 @@ \subsubsection{MAP request}\label{sec:Device Types / IOMMU Device / Device opera
> >  
> >  The driver SHOULD set undefined \field{flags} bits to zero.
> >  
> > +The driver SHOULD NOT send MAP requests on a bypass domain.
> > +
> >  \field{virt_end} MUST be strictly greater than \field{virt_start}.
> >  
> >  The driver SHOULD set the VIRTIO_IOMMU_MAP_F_MMIO flag when the physical
> > @@ -538,6 +573,8 @@ \subsubsection{UNMAP request}\label{sec:Device Types / IOMMU Device / Device ope
> >  or be outside any mapping. The last address of a range MUST either be the
> >  last address of a mapping or be outside any mapping.
> >  
> > +The driver SHOULD NOT send UNMAP requests on a bypass domain.
> > +
> >  \devicenormative{\paragraph}{UNMAP request}{Device Types / IOMMU Device / Device operations / UNMAP request}
> >  
> >  If the \field{reserved} field of an UNMAP request is not zero, the device
> Shouldn't we descrive the behavior of the device is MAP/UNMAP is called
> on a bypass domain or if we try to attach another device to a domain
> with a different ATTACH_F_BYPASS flag?

Ok

Thanks,
Jean