On Mon, 2021-09-13 at 20:35 +0200, Paolo Bonzini wrote: > On 13/09/21 17:29, Dave Hansen wrote: > > On 9/13/21 8:14 AM, Paolo Bonzini wrote: > > > On 13/09/21 16:55, Dave Hansen wrote: > > > > > By "Windows startup" I mean even after guest reboot. Because another > > > > > process could sneak in and steal your EPC pages between a close() and an > > > > > open(), I'd like to have a way to EREMOVE the pages while keeping them > > > > > assigned to the specific vEPC instance, i.e.*without* going through > > > > > sgx_vepc_free_page(). > > > > Oh, so you want fresh EPC state for the guest, but you're concerned that > > > > the previous guest might have left them in a bad state. The current > > > > method of getting a new vepc instance (which guarantees fresh state) has > > > > some other downsides. > > > > > > > > Can't another process steal pages via sgxd and reclaim at any time? > > > > > > vEPC pages never call sgx_mark_page_reclaimable, don't they? > > > > Oh, I was just looking that they were on the SGX LRU. You might be right. > > But, we certainly don't want the fact that they are unreclaimable today > > to be part of the ABI. It's more of a bug than a feature. > > Sure, that's fine. > > > > > What's the extra concern here about going through a close()/open() > > > > cycle? Performance? > > > > > > Apart from reclaiming, /dev/sgx_vepc might disappear between the first > > > open() and subsequent ones. > > > > Aside from it being rm'd, I don't think that's possible. > > > > Being rm'd would be a possibility in principle, and it would be ugly for > it to cause issues on running VMs. Also I'd like for it to be able to > pass /dev/sgx_vepc in via a file descriptor, and run QEMU in a chroot or > a mount namespace. Alternatively, with seccomp it may be possible to > sandbox a running QEMU process in such a way that open() is forbidden at > runtime (all hotplug is done via file descriptor passing); it is not yet > possible, but it is a goal. AFAIK, as long you have open files for a device, they work as expected. /Jarkko
