> From: Jason Gunthorpe <jgg@xxxxxxxxxx>
> Sent: Monday, July 31, 2023 9:46 PM
>
> On Mon, Jul 31, 2023 at 08:33:55AM +0000, Liu, Yi L wrote:
> > > From: Jason Gunthorpe <jgg@xxxxxxxxxx>
> > > Sent: Thursday, July 27, 2023 10:43 PM
> > >
> > > On Thu, Jul 27, 2023 at 07:57:57AM +0000, Tian, Kevin wrote:
> > > > > From: Liu, Yi L <yi.l.liu@xxxxxxxxx>
> > > > > Sent: Monday, July 24, 2023 7:00 PM
> > > > >
> > > > > @@ -252,11 +258,20 @@ struct iommu_iotlb_gather {
> > > > > * @remove_dev_pasid: Remove any translation configurations of a specific
> > > > > * pasid, so that any DMA transactions with this pasid
> > > > > * will be blocked by the hardware.
> > > > > + * @hw_info_type: One of enum iommu_hw_info_type defined in
> > > > > + * include/uapi/linux/iommufd.h. It is used to tag the type
> > > > > + * of data returned by .hw_info callback. The drivers that
> > > > > + * support .hw_info callback should define a unique type
> > > > > + * in include/uapi/linux/iommufd.h. For the drivers that do
> > > > > + * not implement .hw_info callback, this field is
> > > > > + * IOMMU_HW_INFO_TYPE_NONE which is 0. Hence, such drivers
> > > > > + * do not need to care this field.
> > > >
> > > > every time looking at this field the same question came out why it is required
> > > > (and looks I forgot your previous response).
> > > >
> >
> > The major reason is that not every driver implements the hw_info
> > callback.
> >
> > > > e.g. why cannot the type be returned in @hw_info():
> > > >
> > > > void *(*hw_info)(struct device *dev, u32 *length, int *type);
> > >
> > > u32 *type
> > >
> > > > NULL callback implies IOMMU_HW_INFO_TYPE_NONE.
> > >
> > > If every one of these queries has its own type it makes sense
> > >
> > > Though, is it not possible that we can have a type for the entire
> > > driver?
> >
> > Not quite sure if I got your point. Is it acceptable to define the
> > callabck in the current version? or Kevin's suggestion makes
> > more sense?
>
> I'm trying to remember if there is a reason we need unique types for
> the domain and the invalidation or if we can get bye with a single
> type just for the whole iommu driver.
I see. Seems like your comment is more related to the below patches.
https://lore.kernel.org/linux-iommu/20230724110406.107212-2-yi.l.liu@xxxxxxxxx/
https://lore.kernel.org/linux-iommu/20230724110406.107212-10-yi.l.liu@xxxxxxxxx/
https://lore.kernel.org/linux-iommu/20230724111335.107427-2-yi.l.liu@xxxxxxxxx/
https://lore.kernel.org/linux-iommu/20230724111335.107427-8-yi.l.liu@xxxxxxxxx/
I think we unique types fort the domain and invalidation.
E.g. IOMMU_HWPT_TYPE_VTD_S1. The reason is that different vendors have
different stage1 format, and require different user parameters to allocate.
So needs to define unique types.
> I suppose if we ever want to to "virtio-iommu invalidation" we'd want
> to use a new type for it?
Yes. needed in the domain allocation path as well. IIRC, there was a
discussion on whether have a general cache invalidation data structure
or not[1], and the conclusion was to have separate invalidation data
structures instead of a generic structure for all types of stage1 page tables.
[1] https://lore.kernel.org/linux-iommu/20230309134217.GA1673607@myrica/
Regards,
Yi Liu
