Re: [PATCH v6 2/6] iommufd: Add IOMMU_HWPT_INVALIDATE

[Jason Gunthorpe <jgg@nvidia.com>]

On Thu, Dec 07, 2023 at 09:04:00AM +0000, Tian, Kevin wrote:
> > From: Liu, Yi L <yi.l.liu@intel.com>
> > Sent: Thursday, December 7, 2023 2:59 PM
> > 
> > On 2023/11/17 21:07, Yi Liu wrote:
> > > @@ -613,4 +614,38 @@ struct iommu_hwpt_get_dirty_bitmap {
> > >   #define IOMMU_HWPT_GET_DIRTY_BITMAP _IO(IOMMUFD_TYPE, \
> > >
> > 	IOMMUFD_CMD_HWPT_GET_DIRTY_BITMAP)
> > >
> > > +/**
> > > + * struct iommu_hwpt_invalidate - ioctl(IOMMU_HWPT_INVALIDATE)
> > > + * @size: sizeof(struct iommu_hwpt_invalidate)
> > > + * @hwpt_id: HWPT ID of a nested HWPT for cache invalidation
> > > + * @reqs_uptr: User pointer to an array having @req_num of cache
> > invalidation
> > > + *             requests. The request entries in the array are of fixed width
> > > + *             @req_len, and contain a user data structure for invalidation
> > > + *             request specific to the given hardware page table.
> > > + * @req_type: One of enum iommu_hwpt_data_type, defining the data
> > type of all
> > > + *            the entries in the invalidation request array. It should suit
> > > + *            with the data_type passed per the allocation of the hwpt pointed
> > > + *            by @hwpt_id.
> > 
> > @Jason and Kevin,
> > 
> > Here a check with you two. I had a conversation with Nic on the definition
> > of req_type here. It was added to support potential multiple kinds of cache
> > invalidation data types for a invalidating cache for a single hwpt type[1].
> > But we defined it as reusing the hwpt_data_type. In this way, it is not
> > able to support the potential case in[1]. is it? Shall we define a separate
> > enum for invalidation data types? And how can we let user know the
> > available invalidation data types for a hwpt type? Any idea?
> > 
> > [1] https://lore.kernel.org/linux-
> > iommu/20231018163720.GA3952@nvidia.com/
> > 
> 
> From that thread Jason mentioned to make the invalidation format
> part of domain allocation. If that is the direction to go then there
> won't be multiple invalidation formats per hwpt. The user should
> create multiple hwpt's per invalidation format (though mixing
> formats in one virtual platform is very unlikely)?

I think we could do either, but I have a vauge cleanness preference
that the enums are just different? That would follow a pretty typical
pattern for a structure tag to reflect the content of the structure.

Jason