Re: [PATCH v17 07/10] mm: introduce memfd_secret system call to create "secret" memory areas

[James Bottomley]

On Tue, 2021-02-16 at 17:34 +0100, David Hildenbrand wrote:
> On 16.02.21 17:25, James Bottomley wrote:
> > On Mon, 2021-02-15 at 20:20 +0100, Michal Hocko wrote:
> > [...]
> > > > >    What kind of flags are we talking about and why would that
> > > > > be a problem with memfd_create interface? Could you be more
> > > > > specific please?
> > > > 
> > > > You mean what were the ioctl flags in the patch series linked
> > > > above? They were SECRETMEM_EXCLUSIVE and SECRETMEM_UNCACHED in
> > > > patch 3/5.
> > > 
> > > OK I see. How many potential modes are we talking about? A few or
> > > potentially many?
> >   
> > Well I initially thought there were two (uncached or not) until you
> > came up with the migratable or non-migratable, which affects the
> > security properties.  But now there's also potential for hardware
> > backing, like mktme,  described by flags as well.  I suppose you
> > could also use RDT to restrict which cache the data goes into: say
> > L1 but not L2 on to lessen the impact of fully uncached (although
> > the big thrust of uncached was to blunt hyperthread side
> > channels).  So there is potential for quite a large expansion even
> > though I'd be willing to bet that a lot of the modes people have
> > thought about turn out not to be very effective in the field.
> 
> Thanks for the insight. I remember that even the "uncached" parts
> was effectively nacked by x86 maintainers (I might be wrong).

It wasn't liked by x86 maintainers, no.  Plus there's no
architecturally standard mechanism for making a page uncached and, as
the arm people pointed out, sometimes no way of ensuring it's never
cached.

>  For the other parts, the question is what we actually want to let
> user space configure.
> 
> Being able to specify "Very secure" "maximum secure" "average
> secure"  all doesn't really make sense to me.

Well, it doesn't to me either unless the user feels a cost/benefit, so
if max cost $100 per invocation and average cost nothing, most people
would chose average unless they had a very good reason not to.  In your
migratable model, if we had separate limits for non-migratable and
migratable, with non-migratable being set low to prevent exhaustion,
max secure becomes a highly scarce resource, whereas average secure is
abundant then having the choice might make sense.

>  The discussion regarding migratability only really popped up because
> this is a user-visible thing and not being able to migrate can be a
> real problem (fragmentation, ZONE_MOVABLE, ...).

I think the biggest use will potentially come from hardware
acceleration.  If it becomes simple to add say encryption to a secret
page with no cost, then no flag needed.  However, if we only have a
limited number of keys so once we run out no more encrypted memory then
it becomes a costly resource and users might want a choice of being
backed by encryption or not.

James