Xu Yilun wrote:
[..]
> This field is intended for out-of-blob values, like fw_ret. But fw_ret
> is specified in GHCB and is vendor specific. Other vendors may also
> have different values of this kind.
>
> So I intend to gather these out-of-blob values in type_info, like:
>
> enum pci_tsm_guest_req_type {
> PCI_TSM_GUEST_REQ_TDXC,
> PCI_TSM_GUEST_REQ_SEV_SNP,
> };
>
> /* SEV SNP guest request type info */
> struct pci_tsm_guest_req_sev_snp {
> s32 fw_err;
> };
>
> Since IOMMUFD has the userspace entry, maybe these definitions should be
> moved to include/uapi/linux/iommufd.h.
>
> In pci-tsm.h, just define:
>
> struct pci_tsm_guest_req_info {
> u32 type;
> void __user *type_info;
> size_t type_info_len;
> void __user *req;
> size_t req_len;
> void __user *resp;
> size_t resp_len;
> };
>
> BTW: TDX Connect has no out-of-blob value, so should set type_info_len = 0
2 comments:
No, I do not think 'struct pci_tsm_guest_req_info' needs the
sophistication of a "type-info + type-info-len" scheme. Just make it a
64-bit @fw_ret property. The kernel has no need for that value, and
there is nothing to indicate it needs to be larger than 64-bit. Both
ends of the pipe understand what @fw_ret might contain.
Yes, this envelope definition belongs in a uapi header.
