Re: [PATCH v13 2/5] tee: generic TEE subsystem

[Arnd Bergmann]

On Thursday, January 19, 2017 5:45:43 PM CET Jens Wiklander wrote:
> 
> On Wed, Jan 18, 2017 at 09:19:25PM +0100, Arnd Bergmann wrote:
> > On Friday, November 18, 2016 3:51:37 PM CET Jens Wiklander wrote:
> > > Initial patch for generic TEE subsystem.
> > > This subsystem provides:
> > > * Registration/un-registration of TEE drivers.
> > > * Shared memory between normal world and secure world.
> > > * Ioctl interface for interaction with user space.
> > > * Sysfs implementation_id of TEE driver
> > > 
> > > A TEE (Trusted Execution Environment) driver is a driver that interfaces
> > > with a trusted OS running in some secure environment, for example,
> > > TrustZone on ARM cpus, or a separate secure co-processor etc.
> > > 
> > > The TEE subsystem can serve a TEE driver for a Global Platform compliant
> > > TEE, but it's not limited to only Global Platform TEEs.
> > > 
> > > This patch builds on other similar implementations trying to solve
> > > the same problem:
> > > * "optee_linuxdriver" by among others
> > >   Jean-michel DELORME<jean-michel.delorme@xxxxxx> and
> > >   Emmanuel MICHEL <emmanuel.michel@xxxxxx>
> > > * "Generic TrustZone Driver" by Javier González <javier@xxxxxxxxxxx>
> > 
> > Can you give an example for a system that would contain more than one
> > TEE? I see that you support dynamic registration, and it's clear that
> > there can be more than one type of TEE, but why would one have more
> > than one at a time, and why not more than 32?
> 
> I know that ST has systems where there's one TEE in TrustZone and
> another TEE on a separate secure co-processor. If you have several TEEs
> it's probably because they have different capabilities (performance
> versus level of security). Just going beyond two or three different
> levels of security with different TEEs sounds a bit extreme, so a
> maximum of 32 or 16 should be fairly safe. If it turns out I'm wrong in
> this assumption it's not that hard to correct it.

Ok

> > 
> > > +	if (copy_from_user(&arg, uarg, sizeof(arg)))
> > > +		return -EFAULT;
> > > +
> > > +	if (sizeof(arg) + TEE_IOCTL_PARAM_SIZE(arg.num_params) != buf.buf_len)
> > > +		return -EINVAL;
> > > +
> > > +	if (arg.num_params) {
> > > +		params = kcalloc(arg.num_params, sizeof(struct tee_param),
> > > +				 GFP_KERNEL);
> > > +		if (!params)
> > > +			return -ENOMEM;
> > 
> > It would be good to have an upper bound on the number of parameters
> > to limit the size of the memory allocation here.
> 
> This is already limited due to:
> 
> The test with: buf.buf_len > TEE_MAX_ARG_SIZE
> 
> And then another test that the number of parameters matches the buffer size
> with: sizeof(arg) + TEE_IOCTL_PARAM_SIZE(arg.num_params) != buf.buf_len

Ok, makes sense.

> > 
> > > +/**
> > > + * struct tee_ioctl_param - parameter
> > > + * @attr: attributes
> > > + * @memref: a memory reference
> > > + * @value: a value
> > > + *
> > > + * @attr & TEE_PARAM_ATTR_TYPE_MASK indicates if memref or value is used in
> > > + * the union. TEE_PARAM_ATTR_TYPE_VALUE_* indicates value and
> > > + * TEE_PARAM_ATTR_TYPE_MEMREF_* indicates memref. TEE_PARAM_ATTR_TYPE_NONE
> > > + * indicates that none of the members are used.
> > > + */
> > > +struct tee_ioctl_param {
> > > +	__u64 attr;
> > > +	union {
> > > +		struct tee_ioctl_param_memref memref;
> > > +		struct tee_ioctl_param_value value;
> > > +	} u;
> > > +};
> > > +
> > > +#define TEE_IOCTL_UUID_LEN		16
> > > +
> > 
> > Having a union in an ioctl argument seems odd. Have you considered
> > using two different ioctl command numbers depending on the type?
> 
> struct tee_ioctl_param is used as an array and some parameters can be
> memrefs while other are values.

Got it. I still think it's a bit awkward on the user ABI side.
I also see that (unlike the in-kernel interface) tee_ioctl_param_memref
and tee_ioctl_param_value are both defined in terms of three __u64
members.

How about simply using one format here and making this

struct tee_ioctl_param {
   __u64 attr;
   __u64 a;
   __u64 b;
   __u64 c;
};

Given that you need a wrapper to set the pointer in memref anyway?

Having an ioctl with a variable number of variable type arguments
is really a weakness of the ABI, but I don't see a good way around
it either, the above would just make it slightly more direct.

> > > +/**
> > > + * struct tee_iocl_supp_send_arg - Send a response to a received request
> > > + * @ret:	[out] return value
> > > + * @num_params	[in] number of parameters following this struct
> > > + */
> > > +struct tee_iocl_supp_send_arg {
> > > +	__u32 ret;
> > > +	__u32 num_params;
> > > +	/*
> > > +	 * this struct is 8 byte aligned since the 'struct tee_ioctl_param'
> > > +	 * which follows requires 8 byte alignment.
> > > +	 *
> > > +	 * Commented out element used to visualize the layout dynamic part
> > > +	 * of the struct. This field is not available at all if
> > > +	 * num_params == 0.
> > > +	 *
> > > +	 * struct tee_ioctl_param params[num_params];
> > > +	 */
> > > +} __aligned(8);
> > 
> > I'd make that 
> > 
> > 	struct tee_ioctl_param params[0];
> > 
> > as wel here, as I also commented in patch 3 that has a similar structure.
> 
> I'm concerned that this may cause warnings when compiling for user space
> depending on compiler and options. Am I too cautious here?

See https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html

I actually misremembered it and the syntax I listed is GCC specific,
but C99 allows "flexible arrays". I think there is no problem relying
on C99 here, we already rely on C99 features elsewhere in headers.

	Arnd