On Tue, 2021-02-16 at 18:16 +0100, David Hildenbrand wrote:
[...]

quoted

quoted

  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.

Right. But wouldn't HW support with configurable keys etc. need more 
syscall parameters (meaning, even memefd_secret() as it is would not
be sufficient?). I suspect the simplistic flag approach might not
be sufficient. I might be wrong because I have no clue about MKTME
and friends.

The theory I was operating under is key management is automatic and
hidden, but key scarcity can't be, so if you flag requesting hardware
backing then you either get success (the kernel found a key) or failure
(the kernel is out of keys).  If we actually want to specify the key
then we need an extra argument and we *must* have a new system call.

Anyhow, I still think extending memfd_create() might just be good
enough - at least for now.

I really think this is the wrong approach for a user space ABI.  If we
think we'll ever need to move to a separate syscall, we should begin
with one.  The pain of trying to shift userspace from memfd_create to a
new syscall would be enormous.  It's not impossible (see clone3) but
it's a pain we should avoid if we know it's coming.

 Things like HW support might have requirements we don't even know
yet and that we cannot even model in memfd_secret() right now.

This is the annoying problem with our Linux unbreakable ABI policy: we
get to plan when the ABI is introduced for stuff we don't yet even know
about.

James