[PATCH V3 2/2] tee: add OP-TEE driver

[Mark Rutland]

> > > > I'm also very concerned that the interface exposed to userspace is
> > > > hideously low-level. Surely we'd expect kernel-side drivers to be doing
> > > > the bulk of direct communication to the OP-TEE instance? In the lack of
> > > > a provided rationale I don't see why the current messaging interface
> > > > would make sense.
> > > The kernel-side does all the direct communication since there's where
> > > the SMC is done, but one level above most of the communication is
> > > terminated in user space. Loading of Trusted Applications and other file
> > > system access is done in by a helper process in user space,
> > > tee-supplicant. A large part of the OP-TEE message protocol is
> > > transparent to the kernel.
> >
> > So you expect userspace clients rather than kernel drivers plugging into
> > this framework?
> 
> The OP-TEE message protocol is primarily for the OP-TEE driver. Other
> TEE drivers plugging into this framwork may use this protocol too, but I
> guess that most will use their own message protocol.
> 
> Provided that each TEE driver rolls their own protocol I'm expecting one
> counter part in user space for each TEE driver. The user space client
> will know which kind of TEE it's talking to through TEE_IOC_VERSION.

Surely that means you need to have every possible user-space client
present in a given filesystem, and you need to have all of them try to
probe the FW to figure out whether appropriate FW is present? That
sounds somewhat heavyweight.

To me it would seem a lot better to have the minimal drivers in the
kernel that get probed based on information from the FW. The main TEE
driver would query the generic APIs to discover which features are
exposed, then instantiate the relevant set of TEE-specific drivers based
on TEE_IOC_VERSION and friends. To handle a need for userspace
components you could emit uevents as necessary, though I'm still unclear
on what the userspace components would do.

> > > We're trying to not exclude any TEE implementation by having this low
> > > level TEE_IOC_CMD instead of a high level interface. The problem with
> > > the different TEEs is that they are designed differently and we haven't
> > > been able to find a high level interface that suits all. Applications
> > > aren't expected to use TEE_IOC_CMD directly, instead there's supposed to
> > > be a client lib wich wraps the kernel interface and provides a higher
> > > level interface, for instance a Global Platform Client API in the case
> > > of a GP TEE.
> > >
> > > For OP-TEE we're using the same protocol all the way down to user space,
> > > the advantage is that it's only one protocol to keep track of and we
> > > don't need to translate the entire message (we do need to copy it,
> > > excluding the payload) each time the message is passed to the next
> > > memory space. In the presence of a hypervisor we have
> > > user space -> kernel -> hypervisor -> secure world
> > > Unfortunately some fields has a different meaning in user space and
> > > kernel space. I'll address this in the documentation.
> > >
> > > The OP-TEE helper process, tee-supplicant, is specific to only OP-TEE.
> > > Other TEEs uses helper processes too, but what they do depend on the
> > > design of the TEE. As a consequence more or less all TEEs needs
> > > something specific for that particular TEE in user space to be fully
> > > functional.
> >
> > I'm not sure that your proposed kernel/user split is ideal. How does
> > userspace determine the appropriate TEE client to use? What's required
> > in the way of arbitration between clients?
> 
> Each client loops through /dev/tee[0-9]* until it finds a TEE it can
> communicate with, or if the client is looking for a specific TEE until
> it's found.
> 
> TEE_IOC_VERSION is used to tell which kind of TEE the client is talking
> to. For a library that implements Global Platforms TEE Client API I
> imagine that in TEEC_InitializeContext() the lib will detect which TEE
> it's talking to and initialize the TEEC_Context appropriately.
> 
> For clients that doesn't care about Global Platform APIs I guess that
> they will search for a specific TEE and give up if it's not found.

That covers detection, but what about arbitrartion?

What happens when I have multiple clients which want to communicate with
the same TEE simultaneously?

> tee-supplicant is a special case since it's a helper process for the
> TEE. The will likely be one tee-supplicant implementation
> (tee-supplicant-optee, tee-supplicant-xyz, etc) for each TEE that user
> space can support. tee-supplicant is looking for a TEE to connect to
> through /dev/teepriv[0-9]*.
>
> The reason for having /dev/teeX for normal clients and /dev/teeprivX for
> tee-supplicants we'd like to have any easy way to set different permission
> on the devices.

What do TEE supplicants do?

[...]

> > How does having more than one TEE work given there's a single conduit?
> 
> Each TEE gets its own /dev/teeXX when the specific TEE driver registers
> in the framework.

Ok.

> > > > > +/*
> > > > > + * Cache settings for shared memory
> > > > > + */
> > > > > +#define OPTEE_SMC_SHM_NONCACHED                0ULL
> > > > > +#define OPTEE_SMC_SHM_CACHED           1ULL
> > > >
> > > > What precise set of memory attributes do these imply?
> > > OPTEE_SMC_SHM_NONCACHED is generally not used, but supposed to match how
> > > the kernel maps noncached memory. OP-TEE maps this as Device-nGnRE
> > > Outer sharable memory (MAIR ATTR = 0x04)
> > >
> > > OPTEE_SMC_SHM_CACHED is cached memory with settings matching how the
> > > kernel maps cached memory. OP-TEE maps this as as Normal Memory,
> > > Outer Write-back non-transient Outer Read Allocate Outer Write Allocate
> > > Inner Write-back non-transient Inner Read Allocate Inner Write Allocate
> > > Inner sharable (MAIR ATTR = 0xff).
> > >
> > > OP-TEE is more or less always compiled for a specific platform so if the
> > > kernel uses some other mapping for a particular platform we'll change the
> > > OP-TEE settings to be compatible with the kernel on that platform.
> >
> > That assumes that the TEE has to know about any kernel that might run.
> > It also implies that a kernel needs to know what each TEE thinks the
> > kernel will be mapping memory as, so it can work around whatever
> > decision has been made by the TEE.
> >
> > So as it stands I think that's a broken design. The attributes you need
> > should be strictly specified. It's perfectly valid for that strict
> > specification to be the same attributes the kernel uses now, but the
> > spec can't change later.
> >
> > Otherwise mismatched attributes will get in the way on some platform,
> > and it's going to be close to impossible to fix things up.
> 
> OK, I see the problem. Is it OK only specify the attributes that need to
> be compatible like:
> #define OPTEE_SMC_SHM_ICACHED           (1 << 0)
> #define OPTEE_SMC_SHM_IWRITE_THROUGH    (1 << 1)
> #define OPTEE_SMC_SHM_IWRITE_BACK       (1 << 2)
> #define OPTEE_SMC_SHM_ISHARABLE         (1 << 3)
> #define OPTEE_SMC_SHM_OCACHED           (1 << 4)
> #define OPTEE_SMC_SHM_OWRITE_THROUGH    (1 << 5)
> #define OPTEE_SMC_SHM_OWRITE_BACK       (1 << 6)
> #define OPTEE_SMC_SHM_OSHARABLE         (1 << 7)
> 
> #define OPTEE_SMC_SHM_CACHED \
>         (OPTEE_SMC_SHM_ICACHED | OPTEE_SMC_SHM_IWRITE_BACK | \
>          OPTEE_SMC_SHM_ISHARABLE | OPTEE_SMC_SHM_OCACHED | \
>          OPTEE_SMC_SHM_OWRITE_BACK)

I'm not sure I follow the question. Will these specific attributes be
mandated by the OP-TEE spec? The set of attributes above are certainly
better specified than simply "CACHED", though it would be nice to have
an architectural definition rather than just a bag of bits.

The architecture maintainers will need to look at the memory attributes
too. I don't think that current APIs offer fine-grained control over
attributes and a UP kernel may not map memory as shareable, for example.

Thanks,
Mark.