Hi Mark,

Le Fri, 23 Jul 2021 15:43:18 +0100,
Mark Brown [off-list ref] a écrit :

On Fri, Jul 23, 2021 at 03:52:37PM +0200, Clément Léger wrote:

quoted

When running under secure monitor control, some controllers can be
placed in secure world and their access is thus not possible from
normal world. However, these controllers frequently contain
registers than are needed by the normal world for a few specific
operations.  

quoted

This patch adds a regmap where registers are accessed using SMCs.
The secure monitor is then responsible to allow or deny access to
the requested registers.  

I can't see any SMC specification for this interface?  Frankly I have
some very substantial concerns about the use case for this over
exposing the functionality of whatever device the SMC is gating
access to through SMC interfaces specific to that functionality.

This would require to modify drivers to check if the access should be
done using SMCs, parse the device tree to find appropriate SMC ids for
each functionality, add dependencies in KConfig on
HAVE_ARM_SMCCC_DISCOVERY, and do SMC calls instead of regmap access.
I'm not saying this is not the way to go but this is clearly more
intrusive than keeping the existing syscon support.

Exposing raw access to a (presumed?) subset of whatever device
functionality feels like the wrong abstraction level to be working at
and like an invitation to system integrators to do things that are
going to get them into trouble down the line.

Indeed, access is reduced to a subset of registers offset which are
checked by the TEE.

If the end user really is just twiddling a few bits here and there I'd
expect those functionality specific services to be pretty simple to
do, slightly more effort on the secure monitor side but a lot safer.

The SMC id is supposed to be unique for a given device. The TEE check is
merely a register offset check and a value check. But I agree that the
attack surface is larger than with a SMC targeted for a single
functionality though.

If there is a use case for passing through an entire device for some
reason (ran out of controllers or something?) then I think we
probably want an abstraction at the bus level so we don't need to add
custom support to every device that we want to pass through and it's
clear what's going on.

In our use case, only a few registers located in a secure controller
is needed to be done. We don't have a use case for an entire device
access.

Clément