On 09/09/20 14:44, Laszlo Ersek wrote: > To summarize: for QemuFlashFvbServicesRuntimeDxe to allocate UEFI > Runtime Services Data type memory, for its own runtime GHCB, two > permissions are necessary (together), at OS runtime: > > - QemuFlashFvbServicesRuntimeDxe must be allowed to swap MSR_SEV_ES_GHCB > temporarily (before executing VMGEXIT), > > - QemuFlashFvbServicesRuntimeDxe must be allowed to change the OS-owned > PTE temporarily (for remapping the GHCB as plaintext, before writing > to it). Condition#2 gets worse: If the firmware-allocated runtime GHCB happens to be virt-mapped by the OS using a 2MB page (covering other UEFI runtime data areas, perhaps?), then simply flipping the encryption bit in QemuFlashFvbServicesRuntimeDxe would mark more runtime memory than just the GHCB as "plaintext". (2MB-4KB specifically.) This could result in: - firmware read accesses outside of the GHCB returning garbage - firmware write accesses ouside of the GHCB leaking information to the hypervisor, and reading back later as garbage In order to prevent those symptoms, the firmware would have to split the 2MB page to 4KB pages, and decrypt just the one (GHCB) page. But page splitting requires additional memory (for the finer granularity page table), and fw memory cannot be allocated at OS runtime. So that extra memory too would have to be pre-allocated by the firmware. Nasty. I'd recommend sticking with this kernel patch. Thanks, Laszlo _______________________________________________ Virtualization mailing list Virtualization@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linuxfoundation.org/mailman/listinfo/virtualization
