[PATCH v2] drm/amdgpu: Avoid possible buffer overflow

[Su Hui]

On 2023/8/21 17:31, Christian König wrote:
> Am 21.08.23 um 09:37 schrieb Su Hui:
>> smatch error:
>> drivers/gpu/drm/amd/amdgpu/amdgpu_discovery.c:1257 
>> amdgpu_discovery_reg_base_init() error:
>> testing array offset 'adev->vcn.num_vcn_inst' after use.
>>
>> change the assignment order to avoid buffer overflow.
>>
>> Fixes: c40bdfb2ffa4 ("drm/amdgpu: fix incorrect VCN revision in SRIOV")
>> Signed-off-by: Su Hui <suhui at nfschina.com>
>> ---
>> changes in v2:
>>   - fix the error about ip->revision (thanks to Christophe JAILLET).
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_discovery.c | 6 +++---
>>   1 file changed, 3 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_discovery.c 
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_discovery.c
>> index 8e1cfc87122d..b07bfd106a9b 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_discovery.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_discovery.c
>> @@ -1250,11 +1250,10 @@ static int 
>> amdgpu_discovery_reg_base_init(struct amdgpu_device *adev)
>>                    *     0b10 : encode is disabled
>>                    *     0b01 : decode is disabled
>>                    */
>> - adev->vcn.vcn_config[adev->vcn.num_vcn_inst] =
>> -                    ip->revision & 0xc0;
>> -                ip->revision &= ~0xc0;
>>                   if (adev->vcn.num_vcn_inst <
>>                       AMDGPU_MAX_VCN_INSTANCES) {
>> + adev->vcn.vcn_config[adev->vcn.num_vcn_inst] =
>> +                        ip->revision & 0xc0;
>>                       adev->vcn.num_vcn_inst++;
>>                       adev->vcn.inst_mask |=
>>                           (1U << ip->instance_number);
>> @@ -1265,6 +1264,7 @@ static int 
>> amdgpu_discovery_reg_base_init(struct amdgpu_device *adev)
>>                           adev->vcn.num_vcn_inst + 1,
>>                           AMDGPU_MAX_VCN_INSTANCES);
>>                   }
>> +                ip->revision &= ~0xc0;
>
> That doesn't looks correct either. The assignment is intentionally 
> outside of the "if".
>
> See "adev->vcn.vcn_config[adev->vcn.num_vcn_inst] = ip->revision & 
> 0xc0;" is always valid.

Hi,

if "adev->vcn.vcn_config[adev->vcn.num_vcn_inst] = ip->revision & 0xc0;" 
is always valid, then

"adev->vcn.num_vcn_inst< AMDGPU_MAX_VCN_INSTANCES " is always true. So 
the below judgement has

no sense.

                   if (adev->vcn.num_vcn_inst <
                       AMDGPU_MAX_VCN_INSTANCES) {

On the contrary, if we need this judgement, then 
"adev->vcn.vcn_config[adev->vcn.num_vcn_inst] = ip->revision & 0xc0;"is not

always valid, because "adev->vcn.num_vcn_inst >= 
AMDGPU_MAX_VCN_INSTANCES" can be true, which cause buffer overflow.

So I think this patch has some sense if I don't make some mistakes.

Su Hui

>
> We just avoid incrementing num_vcn_inst when we already have to many.
>
> Regards,
> Christian.
>
>
>>               }
>>               if (le16_to_cpu(ip->hw_id) == SDMA0_HWID ||
>>                   le16_to_cpu(ip->hw_id) == SDMA1_HWID ||
>