[PATCH 2/2] drm/amd/amdgpu: skip amdgpu_gem_va_update_vm when shadow page directory is swapped out

[Christian König]

Am 12.12.2016 um 16:04 schrieb Nicolai Hähnle:
> From: Nicolai Hähnle <nicolai.haehnle at amd.com>
>
> This fixes a crash with the following stack trace:
>
>   amdgpu_gtt_mgr_alloc
>   amdgpu_ttm_bind
>   amdgpu_vm_update_page_directory
>   amdgpu_gem_va_update_vm
>
> Signed-off-by: Nicolai Hähnle <nicolai.haehnle at amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 11 +++++++++++
>   1 file changed, 11 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
> index fbfbe5a..ffb3e70 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
> @@ -495,20 +495,21 @@ static int amdgpu_gem_va_check(void *param, struct amdgpu_bo *bo)
>    *
>    * Update the bo_va directly after setting it's address. Errors are not
>    * vital here, so they are not reported back to userspace.
>    */
>   static void amdgpu_gem_va_update_vm(struct amdgpu_device *adev,
>   				    struct amdgpu_bo_va *bo_va,
>   				    uint32_t operation)
>   {
>   	struct ttm_validate_buffer tv, *entry;
>   	struct amdgpu_bo_list_entry vm_pd;
> +	struct amdgpu_bo *pd_shadow;
>   	struct ww_acquire_ctx ticket;
>   	struct list_head list, duplicates;
>   	unsigned domain;
>   	int r;
>   
>   	INIT_LIST_HEAD(&list);
>   	INIT_LIST_HEAD(&duplicates);
>   
>   	tv.bo = &bo_va->bo->tbo;
>   	tv.shared = true;
> @@ -521,20 +522,30 @@ static void amdgpu_gem_va_update_vm(struct amdgpu_device *adev,
>   	if (r)
>   		goto error_print;
>   
>   	list_for_each_entry(entry, &list, head) {
>   		domain = amdgpu_mem_type_to_domain(entry->bo->mem.mem_type);
>   		/* if anything is swapped out don't swap it in here,
>   		   just abort and wait for the next CS */
>   		if (domain == AMDGPU_GEM_DOMAIN_CPU)
>   			goto error_unreserve;

I would just do the check in this loop instead of down below specialized 
for the page directory.

We have only two entries in the list anyway and I actually would suggest 
to have something like a amdgpu_bo_swapped_out() or 
amdgpu_bo_gpu_accessible() helper in amdgpu_object.h.

Regards,
Christian.

>   	}
> +
> +	/* Also abort if the page directory shadow has been swapped out. */
> +	pd_shadow = bo_va->vm->page_directory->shadow;
> +	if (pd_shadow) {
> +		domain = amdgpu_mem_type_to_domain(pd_shadow->tbo.mem.mem_type);
> +
> +		if (domain == AMDGPU_GEM_DOMAIN_CPU)
> +			goto error_unreserve;
> +	}
> +
>   	r = amdgpu_vm_validate_pt_bos(adev, bo_va->vm, amdgpu_gem_va_check,
>   				      NULL);
>   	if (r)
>   		goto error_unreserve;
>   
>   	r = amdgpu_vm_update_page_directory(adev, bo_va->vm);
>   	if (r)
>   		goto error_unreserve;
>   
>   	r = amdgpu_vm_clear_freed(adev, bo_va->vm);