[bug report] drm/amdgpu/gfx7: set USER_SHADER_ARRAY_CONFIG based on disable_cu parameter

[Christian König]

Am 03.08.2016 um 11:16 schrieb Nicolai Hähnle:
> On 03.08.2016 11:09, Dan Carpenter wrote:
>> Hello Nicolai Hähnle,
>>
>> The patch 324c614a819a: "drm/amdgpu/gfx7: set
>> USER_SHADER_ARRAY_CONFIG based on disable_cu parameter" from Jun 17,
>> 2016, leads to the following static checker warning:
>>
>>     drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c:5057 gfx_v7_0_get_cu_info()
>>     error: buffer overflow 'cu_info->bitmap' 4 <= 4
>>
>> drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c
>>   5035  static void gfx_v7_0_get_cu_info(struct amdgpu_device *adev)
>>   5036  {
>>   5037          int i, j, k, counter, active_cu_number = 0;
>>   5038          u32 mask, bitmap, ao_bitmap, ao_cu_mask = 0;
>>   5039          struct amdgpu_cu_info *cu_info = &adev->gfx.cu_info;
>>   5040          unsigned disable_masks[4 * 2];
>>   5041
>>   5042          memset(cu_info, 0, sizeof(*cu_info));
>>   5043
>>   5044          amdgpu_gfx_parse_disable_cu(disable_masks, 4, 2);
>>   5045
>>   5046          mutex_lock(&adev->grbm_idx_mutex);
>>   5047          for (i = 0; i < adev->gfx.config.max_shader_engines; 
>> i++) {
>>   5048                  for (j = 0; j < 
>> adev->gfx.config.max_sh_per_se; j++) {
>>   5049                          mask = 1;
>>   5050                          ao_bitmap = 0;
>>   5051                          counter = 0;
>>   5052                          gfx_v7_0_select_se_sh(adev, i, j, 
>> 0xffffffff);
>>   5053                          if (i < 4 && j < 2)
>>                                     ^^^^^
>> Is it really possible for i to be >= 4?
>
> No, because for that to happen we would have to add support for 
> hardware with > 4 shader engines. What's the idiomatic way to express 
> this kind of assumption in the kernel? 
> BUG_ON(adev->gfx.config.max_shader_engines > 4)? Some other form of 
> assert?

Either WARN_ON() or BUG_ON().

BUG_ON() stops any current processing because we run into such a fatal 
issue that continuing would cause more damage than just stopping the 
current process and waiting forever.

So for this case WARN_ON() sounds more appropriate.

Regards,
Christian.

>
> Nicolai
>
>>
>>   5054 gfx_v7_0_set_user_cu_inactive_bitmap(
>>   5055                                          adev, disable_masks[i 
>> * 2 + j]);
>>   5056                          bitmap = 
>> gfx_v7_0_get_cu_active_bitmap(adev);
>>   5057                          cu_info->bitmap[i][j] = bitmap;
>>                                 ^^^^^^^^^^^^^^^^^^^^^
>> Because if so, then we are screwed here.
>>
>>   5058
>>   5059                          for (k = 0; k < 16; k ++) {
>>   5060                                  if (bitmap & mask) {
>>   5061                                          if (counter < 2)
>>   5062 ao_bitmap |= mask;
>>   5063                                          counter ++;
>>   5064                                  }
>>   5065                                  mask <<= 1;
>>   5066                          }
>>   5067                          active_cu_number += counter;
>>   5068                          ao_cu_mask |= (ao_bitmap << (i * 16 + 
>> j * 8));
>>   5069                  }
>>   5070          }
>>   5071          gfx_v7_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 
>> 0xffffffff);
>>   5072          mutex_unlock(&adev->grbm_idx_mutex);
>>   5073
>>   5074          cu_info->number = active_cu_number;
>>   5075          cu_info->ao_cu_mask = ao_cu_mask;
>>   5076  }
>>
>> regards,
>> dan carpenter
>>
> _______________________________________________
> dri-devel mailing list
> dri-devel at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/dri-devel