[v2,2/2] platform/x86/amd/pmf: Fix possible out-of-bound memory accesses

  The length of the policy buffer is not validated before accessing it,
which means that multiple out-of-bounds memory accesses can occur.

This is especially bad since userspace can load policy binaries over
debugfs.

Compile-tested only.

Signed-off-by: Armin Wolf <W_Armin@gmx.de>
---
Changes since v1:
- check if the policy buffer also has enough room for storing the length
---
 drivers/platform/x86/amd/pmf/tee-if.c | 6 ++++++
 1 file changed, 6 insertions(+)

--
2.39.2
  

Hi Shyam & Armin,

Shyam, please take a look at the question below.

On Tue, 27 Feb 2024, Armin Wolf wrote:

> The length of the policy buffer is not validated before accessing it,
> which means that multiple out-of-bounds memory accesses can occur.
> 
> This is especially bad since userspace can load policy binaries over
> debugfs.

> +	if (dev->policy_sz < POLICY_COOKIE_LEN + sizeof(length))
> +		return -EINVAL;
> +
>  	cookie = *(u32 *)(dev->policy_buf + POLICY_COOKIE_OFFSET);
>  	length = *(u32 *)(dev->policy_buf + POLICY_COOKIE_LEN);

This starts to feel like adding a struct for the header(?) would be better
course of action here as then one could compare against sizeof(*header) 
and avoid all those casts (IMO, just access the header fields directly 
w/o the local variables).

Shyam, do you think a struct makes sense here? There's some header in 
this policy, right?


There are more thing to address here...

1) amd_pmf_start_policy_engine() function returns -EINVAL & res that is 
   TA_PMF_* which inconsistent in type of the return value

2) Once 1) is fixed, the caller shadowing the return code can be fixed as 
   well:
        ret = amd_pmf_start_policy_engine(dev);
        if (ret)
                return -EINVAL;
  

Hi Ilpo,

On 2/27/2024 21:15, Ilpo Järvinen wrote:
> Hi Shyam & Armin,
> 
> Shyam, please take a look at the question below.
> 
> On Tue, 27 Feb 2024, Armin Wolf wrote:
> 
>> The length of the policy buffer is not validated before accessing it,
>> which means that multiple out-of-bounds memory accesses can occur.
>>
>> This is especially bad since userspace can load policy binaries over
>> debugfs.
> 

IMO, this patch is not required, reason being:
- the debugfs patch gets created only when CONFIG_AMD_PMF_DEBUG is
enabled.
- Sideload of policy binaries is only supported with a valid signing
key. (think like this can be tested & verified within AMD environment)
- Also, in amd_pmf_get_pb_data() there are boundary conditions that
are being checked. Is that not sufficient enough?

>> +	if (dev->policy_sz < POLICY_COOKIE_LEN + sizeof(length))
>> +		return -EINVAL;
>> +
>>  	cookie = *(u32 *)(dev->policy_buf + POLICY_COOKIE_OFFSET);
>>  	length = *(u32 *)(dev->policy_buf + POLICY_COOKIE_LEN);
> 
> This starts to feel like adding a struct for the header(?) would be better
> course of action here as then one could compare against sizeof(*header) 
> and avoid all those casts (IMO, just access the header fields directly 
> w/o the local variables).
Not sure if I get your question clearly. Can you elaborate a bit more
on the struct you are envisioning?

but IHMO, we actually don't need a struct - as all that we would need
is to make sure the signing cookie is part of the policy binary and
leave the rest of the error handling to ASP/TEE modules (we can rely
on the feedback from those modules).

> 
> Shyam, do you think a struct makes sense here? There's some header in 
> this policy, right?

Yes, the policy binary on a whole has multiple sections within it and
there are multiple headers (like signing, OEM header, etc).

But that might be not real interest to the PMF driver. The only thing
the driver has to make sure is that the policy binary going into ASP
(AMD Secure Processor) is with in the limits and has a valid signing
cookie. So this part is already taken care in the current code.

> 
> 
> There are more thing to address here...
> 
> 1) amd_pmf_start_policy_engine() function returns -EINVAL & res that is 
>    TA_PMF_* which inconsistent in type of the return value
> 

ah! it has mix of both int and u32 :-)

Armin, would you like to amend this in your current series? or else I
will submit a change for this in my next series.

Thanks,
Shyam

> 2) Once 1) is fixed, the caller shadowing the return code can be fixed as 
>    well:
>         ret = amd_pmf_start_policy_engine(dev);
>         if (ret)
>                 return -EINVAL;
> 
>
  

Am 28.02.24 um 12:16 schrieb Shyam Sundar S K:

> Hi Ilpo,
>
> On 2/27/2024 21:15, Ilpo Järvinen wrote:
>> Hi Shyam & Armin,
>>
>> Shyam, please take a look at the question below.
>>
>> On Tue, 27 Feb 2024, Armin Wolf wrote:
>>
>>> The length of the policy buffer is not validated before accessing it,
>>> which means that multiple out-of-bounds memory accesses can occur.
>>>
>>> This is especially bad since userspace can load policy binaries over
>>> debugfs.
> IMO, this patch is not required, reason being:
> - the debugfs patch gets created only when CONFIG_AMD_PMF_DEBUG is
> enabled.
> - Sideload of policy binaries is only supported with a valid signing
> key. (think like this can be tested & verified within AMD environment)
> - Also, in amd_pmf_get_pb_data() there are boundary conditions that
> are being checked. Is that not sufficient enough?

IMHO, amd_pmf_get_pb_data() only checks if the length of the binary is
between 0 and the maximum buffer size.

If for example the binary contains only 4 bytes, then there will be an
out-of-bounds access when trying to read the cookie and length.

Or if the length is bigger than the binary buffer, then the driver just
updates the buffer length even if the buffer is too small.

I think the driver should catch such cases and return an error.

(Please note that we are talking about the binary buffer, not the internal
structure of the remaining policy binary itself).

>>> +	if (dev->policy_sz < POLICY_COOKIE_LEN + sizeof(length))
>>> +		return -EINVAL;
>>> +
>>>   	cookie = *(u32 *)(dev->policy_buf + POLICY_COOKIE_OFFSET);
>>>   	length = *(u32 *)(dev->policy_buf + POLICY_COOKIE_LEN);
>> This starts to feel like adding a struct for the header(?) would be better
>> course of action here as then one could compare against sizeof(*header)
>> and avoid all those casts (IMO, just access the header fields directly
>> w/o the local variables).
> Not sure if I get your question clearly. Can you elaborate a bit more
> on the struct you are envisioning?

I think he envisions something like this:

struct __packed cookie_header {
	u32 magic;
	u32 length;
};

>
> but IHMO, we actually don't need a struct - as all that we would need
> is to make sure the signing cookie is part of the policy binary and
> leave the rest of the error handling to ASP/TEE modules (we can rely
> on the feedback from those modules).
>
>> Shyam, do you think a struct makes sense here? There's some header in
>> this policy, right?
> Yes, the policy binary on a whole has multiple sections within it and
> there are multiple headers (like signing, OEM header, etc).
>
> But that might be not real interest to the PMF driver. The only thing
> the driver has to make sure is that the policy binary going into ASP
> (AMD Secure Processor) is with in the limits and has a valid signing
> cookie. So this part is already taken care in the current code.
>
>>
>> There are more thing to address here...
>>
>> 1) amd_pmf_start_policy_engine() function returns -EINVAL & res that is
>>     TA_PMF_* which inconsistent in type of the return value
>>
> ah! it has mix of both int and u32 :-)
>
> Armin, would you like to amend this in your current series? or else I
> will submit a change for this in my next series.
>
> Thanks,
> Shyam

I can do so, but i will be unable to send a new patch series for the rest of this week.

Thanks,
Armin Wolf

>> 2) Once 1) is fixed, the caller shadowing the return code can be fixed as
>>     well:
>>          ret = amd_pmf_start_policy_engine(dev);
>>          if (ret)
>>                  return -EINVAL;
>>
>>
  

On Wed, 28 Feb 2024, Armin Wolf wrote:
> Am 28.02.24 um 12:16 schrieb Shyam Sundar S K:
> > On 2/27/2024 21:15, Ilpo Järvinen wrote:
> > > On Tue, 27 Feb 2024, Armin Wolf wrote:
> > > 
> > > > The length of the policy buffer is not validated before accessing it,
> > > > which means that multiple out-of-bounds memory accesses can occur.
> > > > 
> > > > This is especially bad since userspace can load policy binaries over
> > > > debugfs.
> > IMO, this patch is not required, reason being:
> > - the debugfs patch gets created only when CONFIG_AMD_PMF_DEBUG is
> > enabled.
> > - Sideload of policy binaries is only supported with a valid signing
> > key. (think like this can be tested & verified within AMD environment)
> > - Also, in amd_pmf_get_pb_data() there are boundary conditions that
> > are being checked. Is that not sufficient enough?
> 
> IMHO, amd_pmf_get_pb_data() only checks if the length of the binary is
> between 0 and the maximum buffer size.
> 
> If for example the binary contains only 4 bytes, then there will be an
> out-of-bounds access when trying to read the cookie and length.
> 
> Or if the length is bigger than the binary buffer, then the driver just
> updates the buffer length even if the buffer is too small.
> 
> I think the driver should catch such cases and return an error.
> 
> (Please note that we are talking about the binary buffer, not the internal
> structure of the remaining policy binary itself).

Yes. Out of bound accesses are not okay during validation even if the
binary itself would get rejected at a later stage. It doesn't matter if 
the interface is only for debug or wider scope.

> > > > +	if (dev->policy_sz < POLICY_COOKIE_LEN + sizeof(length))
> > > > +		return -EINVAL;
> > > > +
> > > >   	cookie = *(u32 *)(dev->policy_buf + POLICY_COOKIE_OFFSET);
> > > >   	length = *(u32 *)(dev->policy_buf + POLICY_COOKIE_LEN);
> > > This starts to feel like adding a struct for the header(?) would be better
> > > course of action here as then one could compare against sizeof(*header)
> > > and avoid all those casts (IMO, just access the header fields directly
> > > w/o the local variables).
> > Not sure if I get your question clearly. Can you elaborate a bit more
> > on the struct you are envisioning?
> 
> I think he envisions something like this:
> 
> struct __packed cookie_header {
> 	u32 magic;
> 	u32 length;
> };
> 
> > 
> > but IHMO, we actually don't need a struct - as all that we would need
> > is to make sure the signing cookie is part of the policy binary and
> > leave the rest of the error handling to ASP/TEE modules (we can rely
> > on the feedback from those modules).
> > 
> > > Shyam, do you think a struct makes sense here? There's some header in
> > > this policy, right?
> > Yes, the policy binary on a whole has multiple sections within it and
> > there are multiple headers (like signing, OEM header, etc).
> >
> > But that might be not real interest to the PMF driver. The only thing
> > the driver has to make sure is that the policy binary going into ASP
> > (AMD Secure Processor) is with in the limits and has a valid signing
> > cookie. So this part is already taken care in the current code.

Clearly the PMF driver is interested in the header which contains
the cookie and length fields as proven by the code (even if that's for 
the validation purposes only). I'm not asking about add various other 
headers which are of no interest.

> > > There are more thing to address here...
> > > 
> > > 1) amd_pmf_start_policy_engine() function returns -EINVAL & res that is
> > >     TA_PMF_* which inconsistent in type of the return value
> > > 
> > ah! it has mix of both int and u32 :-)

I was talking more on a logical level not so much about C type itself. 

It is confusing to return error in two ways: as -Exx values and 
TA_PMF_* which are positive. As a general rule (IMO), the HW specific 
errors should be mapped to -Exx codes at latest when a function returns 
also -Exx code which is the case here.