On Mon, Jun 17, 2019 at 6:35 PM Kai Huang <kai.huang@xxxxxxxxxxxxxxx> wrote:
>
>
> > > >
> > > > I'm having a hard time imagining that ever working -- wouldn't it blow
> > > > up if someone did:
> > > >
> > > > fd = open("/dev/anything987");
> > > > ptr1 = mmap(fd);
> > > > ptr2 = mmap(fd);
> > > > sys_encrypt(ptr1);
> > > >
> > > > So I think it really has to be:
> > > > fd = open("/dev/anything987");
> > > > ioctl(fd, ENCRYPT_ME);
> > > > mmap(fd);
> > >
> > > This requires "/dev/anything987" to support ENCRYPT_ME ioctl, right?
> > >
> > > So to support NVDIMM (DAX), we need to add ENCRYPT_ME ioctl to DAX?
> >
> > Yes and yes, or we do it with layers -- see below.
> >
> > I don't see how we can credibly avoid this. If we try to do MKTME
> > behind the DAX driver's back, aren't we going to end up with cache
> > coherence problems?
>
> I am not sure whether I understand correctly but how is cache coherence problem related to putting
> MKTME concept to different layers? To make MKTME work with DAX/NVDIMM, I think no matter which layer
> MKTME concept resides, eventually we need to put keyID into PTE which maps to NVDIMM, and kernel
> needs to manage cache coherence for NVDIMM just like for normal memory showed in this series?
>
I mean is that, to avoid cache coherence problems, something has to
prevent user code from mapping the same page with two different key
ids. If the entire MKTME mechanism purely layers on top of DAX,
something needs to prevent the underlying DAX device from being mapped
at the same time as the MKTME-decrypted view. This is obviously
doable, but it's not automatic.
