Re: [Tee-dev] [PATCH v1 00/14] tee: optee: add dynamic shared memory support

[Volodymyr Babchuk]

Hello Stuart,

On 03.10.17 19:05, Stuart Yoder wrote:
> On 9/28/17 1:03 PM, Volodymyr Babchuk wrote:
> > From: Volodymyr Babchuk <vlad.babchuk@xxxxxxxxx>
> >
> > This patch series enables dynamic shared memory support in the TEE
> > subsystem as a whole and in OP-TEE in particular.
> >
> > Global Platform TEE specification [1] allows client applications
> > to register part of own memory as a shared buffer between
> > application and TEE. This allows fast zero-copy communication between
> > TEE and REE. But current implementation of TEE in Linux does not support
> > this feature.
> >
> > Also, current implementation of OP-TEE transport uses fixed size
> > pre-shared buffer for all communications with OP-TEE OS. This is okay
> > in the most use cases. But this prevents use of OP-TEE in virtualized
> > environments, because:
> >   a) We can't share the same buffer between different virtual machines
> >   b) Physically contiguous memory as seen by VM can be non-contiguous
> >      in reality (and as seen by OP-TEE OS) due to second stage of
> >      MMU translation.
> >   c) Size of this pre-shared buffer is limited.
> >
> > So, first part of this patch series adds generic register/unregister
> > interface to tee subsystem. Next patches add necessary features
> > into OP-TEE driver, so it can use not only static pre-shared buffer,
> > but whole RAM to communicate with OP-TEE OS.
>
> As is, the patch series enables dynamic shared memory, but keeps the
> assumption that there must static shared memory as well.  In the case
> of virtual machines this isn't what we want (at least that's the way
> I see KVM working).
Yep, but it is needed for backward compatibility. I can't guarantee, 
that all devices will support dynamic shared memory. Also, hypervisor 
(XEN on KVM) need to intercept all SMCs anyways. My intention was to 
disable static memory buffer in TEE mediator inside hypervisor. 
Hypervisor will trap OPTEE_SMC_GET_SHM_CONFIG call and return error on 
NULL pointers, or something like that. This is my next step.

> The series substantially re-works optee_config_shm_memremap() and it
> seems like it would not be that difficult to remove the assumptions that
> there _must_ be static shared memory available.
Yes, I'm agree. We can remove assumption that there must be static 
shared region. But I think, this is not task for this patch series. This 
patches only add generic feature, that can benefit all clients. I plan 
to add virtualization-specific changes later.
> ...I have some more specific comments on patch 12/14.
>
> Thanks,
> Stuart