On Mon, Aug 18, 2025 at 11:09:44AM +0200, Eugenio Perez Martin wrote:
> On Thu, Aug 14, 2025 at 12:42 PM Michael S. Tsirkin <m...@redhat.com> wrote:
> >
> > On Thu, Aug 14, 2025 at 11:36:22AM +0800, Jason Wang wrote:
> > > > > diff --git a/include/linux/virtio.h b/include/linux/virtio.h
> > > > > index addbc209275a..37029df94aaf 100644
> > > > > --- a/include/linux/virtio.h
> > > > > +++ b/include/linux/virtio.h
> > > > > @@ -40,6 +40,13 @@ struct virtqueue {
> > > > > void *priv;
> > > > > };
> > > > >
> > > > > +union vring_mapping_token {
> > > > > + /* Device that performs DMA */
> > > > > + struct device *dma_dev;
> > > > > + /* Transport specific token used for doing map */
> > > > > + void *opaque;
> > > >
> > > > Please just declare whatever structure you want it to be.
> > >
> > > It's an opaque one and so
> > >
> > > 1) the virtio core knows nothing about that because it could be
> > > transport or device specific
> > > 2) no assumption of the type and usage, it just receive it from the
> > > transport and pass it back when doing the mapping
> > >
> > > It should work like page->private etc.
> > >
> > > Does this make sense?
> > >
> > > Thanks
> >
> > I fully expect most devices simply to use DMA here and no weird
> > tricks. vduse is the weird one, but I don't see us making it
> > grow much beyond that.
> >
> > So I think for now we can just make it vduse_iova_domain *. If we see
> > it's getting out of hand with too many types, we can think of solutions.
> >
>
> I've sent my series of adding ASID to VDUSE, which uses this series'
> token on each vq group, on top of this version of the DMA rework.
But then I see it drops it from vduse_iova_domain. So it does not
look like they union is going to keep growing in an unmanageable way.
> This patch [1] and the next one are the one that reworks the token to
> an empty struct, so virtio can handle it in an opaque way and VDUSE
> can convert it back and forth in a type safe way, skipping the void *.
> Please let me know if you prefer to import a VDUSE header into the
> virtio config header or to make a VDUSE forward declaration instead of
> going through the empty struct to preserve layer boundaries.
Personally for now I'd be happier with just a forward declaration.
I just like seeing things at a glance: "it's a union, and
can be one of two types", is better for me than
"it can be anything grep the source to figure out what it is".
And a forward declaration is opaque by design.
> There is one usage I've not been able to convert though. Jason, could
> you take a look? It is marked as TODO in my series. I'm not sure if
> that's also an abuse of the void * in the DMA rework to be honest, but
> it should be easy to correct.
>
> [1]
> https://lore.kernel.org/all/20250818085711.3461758-4-epere...@redhat.com/T/#u
