On Fri, Aug 15, 2025 at 09:02:09AM +0800, Jason Wang wrote:
> On Thu, Aug 14, 2025 at 6:42 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 *.
>
> It requires exporting VDUSE's internal data structure to the virtio
> layer which is a kind of layer violation.
It's just a forward declaration. You do not need the structure.
> And the iova domain will be
> soon replaced by virtqueue group in Eungeio's series that support
> multiple address spaces
>
> > If we see
> > it's getting out of hand with too many types, we can think of solutions.
>
> Considering we've already at it, I think it's better to have a
> solution now if it's not complicated:
>
> I can think of two kind of solutions:
>
> 1) Stick to the current void * pointer which is simple and straightforward
> 2) Eugenio's suggestion of having a virtio_map_head and let the
> transport/device get the container struct which is the real token.
>
> E.g
>
> struct vduse_iova_domain {
> struct virtio_map_head map;
> ......
> };
>
> If you don't like void *, can we go fo 2)?
>
> Thanks
>
>
> >
> > --
> > MST
> >
