Skip to contentSkip to conversation

drm/amdgpu/gfx6: clean up cu configuration

bug report Dan Carpenter 3 participants 3 messages Mar 31, 2017 – Apr 3, 2017

lore
Share & tools
lore
#0 lore raw

Hello Flora Cui,

The patch 375d6f7057a9: "drm/amdgpu/gfx6: clean up cu configuration" from Feb 7, 2017, leads to the following static checker warning:

	drivers/gpu/drm/amd/amdgpu/gfx_v6_0.c:3737 gfx_v6_0_get_cu_info()
	warn: potential off by one cu_info->bitmap[4]

drivers/gpu/drm/amd/amdgpu/gfx_v6_0.c
3715 static void gfx_v6_0_get_cu_info(struct amdgpu_device *adev) 3716 {

  3717          int i, j, k, counter, active_cu_number = 0;
  3718          u32 mask, bitmap, ao_bitmap, ao_cu_mask = 0;
  3719          struct amdgpu_cu_info *cu_info = &adev->gfx.cu_info;
  3720          unsigned disable_masks[4 * 2];
  3721  
  3722          memset(cu_info, 0, sizeof(*cu_info));
  3723  
  3724          amdgpu_gfx_parse_disable_cu(disable_masks, 4, 2);
  3725  
  3726          mutex_lock(&adev->grbm_idx_mutex);
  3727          for (i = 0; i < adev->gfx.config.max_shader_engines; i++) {
  3728                  for (j = 0; j < adev->gfx.config.max_sh_per_se; j++) {
  3729                          mask = 1;
  3730                          ao_bitmap = 0;
  3731                          counter = 0;
  3732                          gfx_v6_0_select_se_sh(adev, i, j, 0xffffffff);
  3733                          if (i < 4 && j < 2)

If i == 4

  3734                                  gfx_v6_0_set_user_cu_inactive_bitmap(
  3735                                          adev, disable_masks[i * 2 + j]);
  3736                          bitmap = gfx_v6_0_get_cu_enabled(adev);
  3737                          cu_info->bitmap[i][j] = bitmap;

then we are beyond the end of this array. Also, why was this patch even applied when it has no commit message? It's totally not clear to me what the patch is trying to do or why it exists...

  3738  
  3739                          for (k = 0; k < 16; k++) {
  3740                                  if (bitmap & mask) {
  3741                                          if (counter < 2)
  3742                                                  ao_bitmap |= mask;
  3743                                          counter ++;
  3744                                  }
  3745                                  mask <<= 1;
  3746                          }
  3747                          active_cu_number += counter;
  3748                          ao_cu_mask |= (ao_bitmap << (i * 16 + j * 8));
  3749                  }
  3750          }
  3751  
  3752          gfx_v6_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
  3753          mutex_unlock(&adev->grbm_idx_mutex);
  3754  
  3755          cu_info->number = active_cu_number;
  3756          cu_info->ao_cu_mask = ao_cu_mask;
  3757  }

regards,
dan carpenter
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#1 lore raw

Hi Dan Carpenter,

Thank you for the info. This commit is just a clean up to keep align with gfx7/8.

Quoted 31 lines · Dan Carpenter

Hello Flora Cui,

The patch 375d6f7057a9: "drm/amdgpu/gfx6: clean up cu configuration" from Feb 7, 2017, leads to the following static checker warning:

	drivers/gpu/drm/amd/amdgpu/gfx_v6_0.c:3737 gfx_v6_0_get_cu_info()
	warn: potential off by one cu_info->bitmap[4]

drivers/gpu/drm/amd/amdgpu/gfx_v6_0.c
3715 static void gfx_v6_0_get_cu_info(struct amdgpu_device *adev) 3716 {

  3717          int i, j, k, counter, active_cu_number = 0;
  3718          u32 mask, bitmap, ao_bitmap, ao_cu_mask = 0;
  3719          struct amdgpu_cu_info *cu_info = &adev->gfx.cu_info;
  3720          unsigned disable_masks[4 * 2];
  3721  
  3722          memset(cu_info, 0, sizeof(*cu_info));
  3723  
  3724          amdgpu_gfx_parse_disable_cu(disable_masks, 4, 2);
  3725  
  3726          mutex_lock(&adev->grbm_idx_mutex);
  3727          for (i = 0; i < adev->gfx.config.max_shader_engines; i++) {
  3728                  for (j = 0; j < adev->gfx.config.max_sh_per_se; j++) {
  3729                          mask = 1;
  3730                          ao_bitmap = 0;
  3731                          counter = 0;
  3732                          gfx_v6_0_select_se_sh(adev, i, j, 0xffffffff);
  3733                          if (i < 4 && j < 2)

If i == 4

Quoted 8 lines
  3734                                  gfx_v6_0_set_user_cu_inactive_bitmap(
  3735                                          adev, disable_masks[i * 2 + j]);
  3736                          bitmap = gfx_v6_0_get_cu_enabled(adev);
  3737                          cu_info->bitmap[i][j] = bitmap;

then we are beyond the end of this array. Also, why was this patch even applied when it has no commit message? It's totally not clear to me what the patch is trying to do or why it exists...

adev->gfx.config.max_shader_engines is set in gfx_v6_0_gpu_init() and it must be < 4 so we'll never be beyond the end of the array.

Regards,
Flora Cui

Quoted 21 lines
  3738  
  3739                          for (k = 0; k < 16; k++) {
  3740                                  if (bitmap & mask) {
  3741                                          if (counter < 2)
  3742                                                  ao_bitmap |= mask;
  3743                                          counter ++;
  3744                                  }
  3745                                  mask <<= 1;
  3746                          }
  3747                          active_cu_number += counter;
  3748                          ao_cu_mask |= (ao_bitmap << (i * 16 + j * 8));
  3749                  }
  3750          }
  3751  
  3752          gfx_v6_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
  3753          mutex_unlock(&adev->grbm_idx_mutex);
  3754  
  3755          cu_info->number = active_cu_number;
  3756          cu_info->ao_cu_mask = ao_cu_mask;
  3757  }
Quoted 2 lines

regards,
dan carpenter

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#2 lore raw
Quoted 15 lines · Flora Cui
Quoted 15 lines
  3727          for (i = 0; i < adev->gfx.config.max_shader_engines; i++) {
  3728                  for (j = 0; j < adev->gfx.config.max_sh_per_se; j++) {
  3729                          mask = 1;
  3730                          ao_bitmap = 0;
  3731                          counter = 0;
  3732                          gfx_v6_0_select_se_sh(adev, i, j, 0xffffffff);
  3733                          if (i < 4 && j < 2)

If i == 4

  3734                                  gfx_v6_0_set_user_cu_inactive_bitmap(
  3735                                          adev, disable_masks[i * 2 + j]);
  3736                          bitmap = gfx_v6_0_get_cu_enabled(adev);
  3737                          cu_info->bitmap[i][j] = bitmap;
Quoted 5 lines
Quoted 3 lines

then we are beyond the end of this array. Also, why was this patch even applied when it has no commit message? It's totally not clear to me what the patch is trying to do or why it exists...

adev->gfx.config.max_shader_engines is set in gfx_v6_0_gpu_init() and it must be < 4 so we'll never be beyond the end of the array.

Could you remove the condition then? We should either check it always or not at all. We shouldn't check it half the time and assume it's correct the other half.

regards,
dan carpenter

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel