Re: [RFC PATCH v4 20/20] vdpa: Add custom IOTLB translations to SVQ

[Eugenio Perez Martin]

On Wed, Oct 13, 2021 at 7:34 AM Jason Wang <jasowang@redhat.com> wrote:
>
>
> 在 2021/10/1 下午3:06, Eugenio Pérez 写道:
> > Use translations added in VhostIOVATree in SVQ.
> >
> > Now every element needs to store the previous address also, so VirtQueue
> > can consume the elements properly. This adds a little overhead per VQ
> > element, having to allocate more memory to stash them. As a possible
> > optimization, this allocation could be avoided if the descriptor is not
> > a chain but a single one, but this is left undone.
> >
> > TODO: iova range should be queried before, and add logic to fail when
> > GPA is outside of its range and memory listener or svq add it.
> >
> > Signed-off-by: Eugenio Pérez<eperezma@redhat.com>
> > ---
> >   hw/virtio/vhost-shadow-virtqueue.h |   4 +-
> >   hw/virtio/vhost-shadow-virtqueue.c | 130 ++++++++++++++++++++++++-----
> >   hw/virtio/vhost-vdpa.c             |  40 ++++++++-
> >   hw/virtio/trace-events             |   1 +
> >   4 files changed, 152 insertions(+), 23 deletions(-)
>
>
> Think hard about the whole logic. This is safe since qemu memory map
> will fail if guest submits a invalidate IOVA.
>

Can you expand on this? What you mean is that VirtQueue already
protects SVQ code if the guest sets an invalid buffer address (GPA),
isn't it?

> Then I wonder if we do something much more simpler:
>
> 1) Using qemu VA as IOVA but only maps the VA that belongs to guest
> 2) Then we don't need any IOVA tree here, what we need is to just map
> vring and use qemu VA without any translation
>

That would be great, but either qemu's SVQ vring or guest translated
buffers address (in qemu VA form) were already in high addresses,
outside of the device's iova range (in my test).

I didn't try remapping tricks to make them fit in the range, but I
think it does complicate the solution relatively fast if there was
already memory in that range owned by qemu before enabling SVQ:

* Guest memory must be contiguous in VA address space, but it "must"
support hotplug/unplug (although vDPA currently pins it). Hotplug
memory could always overlap with SVQ vring, so we would need to move
it.
* Duplicating mapped memory for writing? (Not sure if guest memory is
actually movable in qemu).
* Indirect descriptors will need to allocate and free memory more or
less frequently, increasing the possibility of overlapping.

If we can move guest memory, however, I can see how we can track it in
a tree *but* mark when the tree is 1:1 with qemu's VA, so buffers
forwarding does not take the translation penalty. When guest memory
cannot be map 1:1, we can resort to tree, and come back to 1:1
translation if the offending tree node(s) get deleted.

However I think this puts the solution a little bit farther than
"starting simple" :).

Does it make sense?

Thanks!

> Thanks
>