On 11/16/26 04:22, Jesse Zhang wrote:
> amdgpu_trap_validate_user_addr() passes the byte address of the second
> level TBA and TMA to amdgpu_vm_bo_lookup_mapping(), but the VM mappings
> are indexed by GPU page number, like every other caller of it does and
> like the range check right after it assumes. The lookup never finds the
> mapping, so SET_L2_TRAP fails with -EINVAL for any valid TBA/TMA.
> 
> On navi48 with the VM ioctl SET_L2_TRAP added, the IGT slth-* tests map
> the TBA at 0x100000000 and the lookup of 0x100000000 returns NULL, while
> the mapping is at page 0x100000. With the page number the lookup finds
> the mapping and slth-{gfx,comp}-kq-vgpr-proof, slth-gfx-umq-vgpr-proof
> and slth-comp-kq-returnproof pass.
> 
> Look the mapping up by page number.
> 
> CC:[email protected]
> CC:Christian Koenig <[email protected]>
> CC:Lijo Lazar <[email protected]>
> CC:Srinivasan Shanmugam <[email protected]>
> 
> Fixes: 1a6f95d4a5d7 ("drm/amdgpu: Add second-level trap handler 
> infrastructure")
> Signed-off-by: Jesse Zhang <[email protected]>
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_trap.c | 13 +++++--------
>  1 file changed, 5 insertions(+), 8 deletions(-)
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_trap.c 
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_trap.c
> index 3660f94a0929..52e4d30e63fa 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_trap.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_trap.c
> @@ -622,17 +622,14 @@ static int amdgpu_trap_validate_user_addr(struct 
> amdgpu_device *adev,
>       if (check_add_overflow(start, (u64)usr_addr->size - 1, &end))
>               return -EINVAL;
>  
> -     /*
> -      * amdgpu_vm_bo_lookup_mapping() expects a GPU VA,
> -      * not a page-shifted address.
> -      */
> -     va_map = amdgpu_vm_bo_lookup_mapping(vm, start);
> -     if (!va_map)
> -             return -EINVAL;
> -
>       addr = start >> AMDGPU_GPU_PAGE_SHIFT;
>       last = end >> AMDGPU_GPU_PAGE_SHIFT;
>  
> +     /* The VM mappings are indexed by GPU page number */
> +     va_map = amdgpu_vm_bo_lookup_mapping(vm, addr);
> +     if (!va_map)
> +             return -EINVAL;
> +
>       if (addr < va_map->start || last > va_map->last)

Only a nit but the addr < va_map->start check here doesn't make sense, only the 
last > va_map->last does.

With that fixed Reviewed-by: Christian König <[email protected]>

Regards,
Christian.

>               return -EINVAL;
>  

Reply via email to