kvm PCI assignment & VFIO ramblings

[Alex Williamson]

On Tue, 2011-08-23 at 10:33 -0700, Aaron Fabbri wrote:
> 
> 
> On 8/23/11 10:01 AM, "Alex Williamson" <alex.williamson at redhat.com> wrote:
> 
> > On Tue, 2011-08-23 at 16:54 +1000, Benjamin Herrenschmidt wrote:
> >> On Mon, 2011-08-22 at 17:52 -0700, aafabbri wrote:
> >> 
> >>> I'm not following you.
> >>> 
> >>> You have to enforce group/iommu domain assignment whether you have the
> >>> existing uiommu API, or if you change it to your proposed
> >>> ioctl(inherit_iommu) API.
> >>> 
> >>> The only change needed to VFIO here should be to make uiommu fd assignment
> >>> happen on the groups instead of on device fds.  That operation fails or
> >>> succeeds according to the group semantics (all-or-none assignment/same
> >>> uiommu).
> >> 
> >> Ok, so I missed that part where you change uiommu to operate on group
> >> fd's rather than device fd's, my apologies if you actually wrote that
> >> down :-) It might be obvious ... bare with me I just flew back from the
> >> US and I am badly jet lagged ...
> > 
> > I missed it too, the model I'm proposing entirely removes the uiommu
> > concept.
> > 
> >> So I see what you mean, however...
> >> 
> >>> I think the question is: do we force 1:1 iommu/group mapping, or do we allow
> >>> arbitrary mapping (satisfying group constraints) as we do today.
> >>> 
> >>> I'm saying I'm an existing user who wants the arbitrary iommu/group mapping
> >>> ability and definitely think the uiommu approach is cleaner than the
> >>> ioctl(inherit_iommu) approach.  We considered that approach before but it
> >>> seemed less clean so we went with the explicit uiommu context.
> >> 
> >> Possibly, the question that interest me the most is what interface will
> >> KVM end up using. I'm also not terribly fan with the (perceived)
> >> discrepancy between using uiommu to create groups but using the group fd
> >> to actually do the mappings, at least if that is still the plan.
> > 
> > Current code: uiommu creates the domain, we bind a vfio device to that
> > domain via a SET_UIOMMU_DOMAIN ioctl on the vfio device, then do
> > mappings via MAP_DMA on the vfio device (affecting all the vfio devices
> > bound to the domain)
> > 
> > My current proposal: "groups" are predefined.  groups ~= iommu domain.
> 
> This is my main objection.  I'd rather not lose the ability to have multiple
> devices (which are all predefined as singleton groups on x86 w/o PCI
> bridges) share IOMMU resources.  Otherwise, 20 devices sharing buffers would
> require 20x the IOMMU/ioTLB resources.  KVM doesn't care about this case?

We do care, I just wasn't prioritizing it as heavily since I think the
typical model is probably closer to 1 device per guest.

> > The iommu domain would probably be allocated when the first device is
> > bound to vfio.  As each device is bound, it gets attached to the group.
> > DMAs are done via an ioctl on the group.
> > 
> > I think group + uiommu leads to effectively reliving most of the
> > problems with the current code.  The only benefit is the group
> > assignment to enforce hardware restrictions.  We still have the problem
> > that uiommu open() = iommu_domain_alloc(), whose properties are
> > meaningless without attached devices (groups).  Which I think leads to
> > the same awkward model of attaching groups to define the domain, then we
> > end up doing mappings via the group to enforce ordering.
> 
> Is there a better way to allow groups to share an IOMMU domain?
> 
> Maybe, instead of having an ioctl to allow a group A to inherit the same
> iommu domain as group B, we could have an ioctl to fully merge two groups
> (could be what Ben was thinking):
> 
> A.ioctl(MERGE_TO_GROUP, B)
> 
> The group A now goes away and its devices join group B.  If A ever had an
> iommu domain assigned (and buffers mapped?) we fail.
> 
> Groups cannot get smaller (they are defined as minimum granularity of an
> IOMMU, initially).  They can get bigger if you want to share IOMMU
> resources, though.
> 
> Any downsides to this approach?

That's sort of the way I'm picturing it.  When groups are bound
together, they effectively form a pool, where all the groups are peers.
When the MERGE/BIND ioctl is called on group A and passed the group B
fd, A can check compatibility of the domain associated with B, unbind
devices from the B domain and attach them to the A domain.  The B domain
would then be freed and it would bump the refcnt on the A domain.  If we
need to remove A from the pool, we call UNMERGE/UNBIND on B with the A
fd, it will remove the A devices from the shared object, disassociate A
with the shared object, re-alloc a domain for A and rebind A devices to
that domain. 

This is where it seems like it might be helpful to make a GET_IOMMU_FD
ioctl so that an iommu object is ubiquitous and persistent across the
pool.  Operations on any group fd work on the pool as a whole.  Thanks,

Alex