kvm PCI assignment & VFIO ramblings

[Alex Williamson]

On Tue, 2011-08-23 at 07:01 +1000, Benjamin Herrenschmidt wrote:
> On Mon, 2011-08-22 at 09:45 -0600, Alex Williamson wrote:
> 
> > Yes, that's the idea.  An open question I have towards the configuration
> > side is whether we might add iommu driver specific options to the
> > groups.  For instance on x86 where we typically have B:D.F granularity,
> > should we have an option not to trust multi-function devices and use a
> > B:D granularity for grouping?
> 
> Or even B or range of busses... if you want to enforce strict isolation
> you really can't trust anything below a bus level :-)
> 
> > Right, we can also combine models.  Binding a device to vfio
> > creates /dev/vfio$GROUP, which only allows a subset of ioctls and no
> > device access until all the group devices are also bound.  I think
> > the /dev/vfio/$GROUP might help provide an enumeration interface as well
> > though, which could be useful.
> 
> Could be tho in what form ? returning sysfs pathes ?

I'm at a loss there, please suggest.  I think we need an ioctl that
returns some kind of array of devices within the group and another that
maybe takes an index from that array and returns an fd for that device.
A sysfs path string might be a reasonable array element, but it sounds
like a pain to work with.

> > 1:1 group<->process is probably too strong.  Not allowing concurrent
> > open()s on the group file enforces a single userspace entity is
> > responsible for that group.  Device fds can be passed to other
> > processes, but only retrieved via the group fd.  I suppose we could even
> > branch off the dma interface into a different fd, but it seems like we
> > would logically want to serialize dma mappings at each iommu group
> > anyway.  I'm open to alternatives, this just seemed an easy way to do
> > it.  Restricting on UID implies that we require isolated qemu instances
> > to run as different UIDs.  I know that's a goal, but I don't know if we
> > want to make it an assumption in the group security model.
> 
> 1:1 process has the advantage of linking to an -mm which makes the whole
> mmu notifier business doable. How do you want to track down mappings and
> do the second level translation in the case of explicit map/unmap (like
> on power) if you are not tied to an mm_struct ?

Right, I threw away the mmu notifier code that was originally part of
vfio because we can't do anything useful with it yet on x86.  I
definitely don't want to prevent it where it makes sense though.  Maybe
we just record current->mm on open and restrict subsequent opens to the
same.

> > Yes.  I'm not sure there's a good ROI to prioritize that model.  We have
> > to assume >1 device per guest is a typical model and that the iotlb is
> > large enough that we might improve thrashing to see both a resource and
> > performance benefit from it.  I'm open to suggestions for how we could
> > include it though.
> 
> Sharing may or may not be possible depending on setups so yes, it's a
> bit tricky.
> 
> My preference is to have a static interface (and that's actually where
> your pet netlink might make some sense :-) to create "synthetic" groups
> made of other groups if the arch allows it. But that might not be the
> best approach. In another email I also proposed an option for a group to
> "capture" another one...

I already made some comments on this in a different thread, so I won't
repeat here.

> > > If that's
> > > not what you're saying, how would the domains - now made up of a
> > > user's selection of groups, rather than individual devices - be
> > > configured?
> > > 
> > > > Hope that captures it, feel free to jump in with corrections and
> > > > suggestions.  Thanks,
> > > 
> 
> Another aspect I don't see discussed is how we represent these things to
> the guest.
> 
> On Power for example, I have a requirement that a given iommu domain is
> represented by a single dma window property in the device-tree. What
> that means is that that property needs to be either in the node of the
> device itself if there's only one device in the group or in a parent
> node (ie a bridge or host bridge) if there are multiple devices.
> 
> Now I do -not- want to go down the path of simulating P2P bridges,
> besides we'll quickly run out of bus numbers if we go there.
> 
> For us the most simple and logical approach (which is also what pHyp
> uses and what Linux handles well) is really to expose a given PCI host
> bridge per group to the guest. Believe it or not, it makes things
> easier :-)

I'm all for easier.  Why does exposing the bridge use less bus numbers
than emulating a bridge?

On x86, I want to maintain that our default assignment is at the device
level.  A user should be able to pick single or multiple devices from
across several groups and have them all show up as individual,
hotpluggable devices on bus 0 in the guest.  Not surprisingly, we've
also seen cases where users try to attach a bridge to the guest,
assuming they'll get all the devices below the bridge, so I'd be in
favor of making this "just work" if possible too, though we may have to
prevent hotplug of those.

Given the device requirement on x86 and since everything is a PCI device
on x86, I'd like to keep a qemu command line something like -device
vfio,host=00:19.0.  I assume that some of the iommu properties, such as
dma window size/address, will be query-able through an architecture
specific (or general if possible) ioctl on the vfio group fd.  I hope
that will help the specification, but I don't fully understand what all
remains.  Thanks,

Alex