On 3/25/26 12:58, Srinivasan Shanmugam wrote:
> PRT (Page Request Table) mappings are not backed by a real buffer.  In

PRT (Partial Resident Texture).

> this case, bo_va is valid, but bo_va->bo is NULL, meaning the mapping
> exists but does not point to any real buffer object.
> 
> amdgpu_gem_va_ioctl() currently mixes CLEAR and PRT handling, which can
> result in incorrect bo_va selection. CLEAR should use bo_va = NULL,
> while PRT should use the special fpriv->prt_va mapping.
> 
> Fix this by clearly selecting bo_va:
> - use fpriv->prt_va for PRT
> - use NULL only for CLEAR
> - use amdgpu_vm_bo_find() for normal BO mappings
> 
> Also, amdgpu_gem_va_update_vm() accesses bo_va->base.bo without checking
> if it is NULL. This is not valid for PRT mappings.
> 
> This keeps CLEAR, PRT, and normal cases separate and avoids invalid
> memory access.
> 
> Cc: Alex Deucher <[email protected]>
> Suggested-by: Christian König <[email protected]>
> Signed-off-by: Srinivasan Shanmugam <[email protected]>
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 18 ++++++++++++++----
>  1 file changed, 14 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c 
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
> index b0ba2bdaf43a..289d6b58b579 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
> @@ -772,8 +772,10 @@ amdgpu_gem_va_update_vm(struct amdgpu_device *adev,
>       if (r)
>               goto error;
>  
> +     /* Only do BO-specific handling if this VA is backed by a real BO */
>       if ((operation == AMDGPU_VA_OP_MAP ||
>            operation == AMDGPU_VA_OP_REPLACE) &&
> +         bo_va->base.bo &&

That is not correct. This branch here should also be taken for PRT mappings.

>           !amdgpu_vm_is_bo_always_valid(vm, bo_va->base.bo)) {
>  
>               /*
> @@ -909,15 +911,23 @@ int amdgpu_gem_va_ioctl(struct drm_device *dev, void 
> *data,
>                       goto error;
>       }
>  
> -     /* Resolve the BO-VA mapping for this VM/BO combination. */
> -     if (abo) {
> +     /* Resolve the BO-VA mapping for this VM/BO combination.
> +      *
> +      * Depending on the case decide bo_va:
> +      * - PRT: use special per-file prt_va (bo_va valid, but bo_va->bo == 
> NULL)
> +      * - CLEAR: no BO involved → bo_va = NULL
> +      * - Normal BO path: lookup mapping from VM
> +      */
> +     if (args->flags & AMDGPU_VM_PAGE_PRT) {
> +             bo_va = fpriv->prt_va;
> +     } else if (args->operation == AMDGPU_VA_OP_CLEAR) {
> +             bo_va = NULL;
> +     } else if (abo) {
>               bo_va = amdgpu_vm_bo_find(&fpriv->vm, abo);
>               if (!bo_va) {
>                       r = -ENOENT;
>                       goto error;
>               }
> -     } else if (args->operation != AMDGPU_VA_OP_CLEAR) {
> -             bo_va = fpriv->prt_va;

That code already looks correct to me. I don't think we need to change anything 
here.

Where is your crash actually coming from?

Regards,
Christian.

>       } else {
>               bo_va = NULL;
>       }

Reply via email to