Re: [PATCH 03/15] dma-buf & drm/amdgpu: remove dma_resv workaround

["Christian König" <christian.koenig@amd.com>]

Am 21.04.22 um 23:13 schrieb Zack Rusin:
> On Thu, 2022-04-21 at 12:17 +0200, Christian König wrote:
>> ⚠ External Email
>>
>> Am 20.04.22 um 21:28 schrieb Zack Rusin:
>>> [SNIP]
>>>> To figure out what it is could you try the following code
>>>> fragment:
>>>>
>>>> diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_validation.c
>>>> b/drivers/gpu/drm/vmwgfx/vmwgfx_validation.c
>>>> index f46891012be3..a36f89d3f36d 100644
>>>> --- a/drivers/gpu/drm/vmwgfx/vmwgfx_validation.c
>>>> +++ b/drivers/gpu/drm/vmwgfx/vmwgfx_validation.c
>>>> @@ -288,7 +288,7 @@ int vmw_validation_add_bo(struct
>>>> vmw_validation_context *ctx,
>>>>                    val_buf->bo = ttm_bo_get_unless_zero(&vbo-
>>>>> base);
>>>>                    if (!val_buf->bo)
>>>>                            return -ESRCH;
>>>> -               val_buf->num_shared = 0;
>>>> +               val_buf->num_shared = 16;
>>>>                    list_add_tail(&val_buf->head, &ctx->bo_list);
>>>>                    bo_node->as_mob = as_mob;
>>>>                    bo_node->cpu_blit = cpu_blit;
>>> Fails the same BUG_ON with num_fences and max_fences == 0.
>> Thanks for testing this.
>>
>> So the buffer object is not reserved through
>> vmw_validation_bo_reserve(), but comes from somewhere else.
>> Unfortunately I absolutely can't find where that's coming from.
>>
>> Do you have some documentation howto setup vmwgfx? E.g. sample VM
>> which
>> I can download somewhere etc..
> I don't have an external machine to upload it to. Getting an external
> machine to run Mesa CI on has been on our todo for a while, so I'll try
> to setup something next week.
>
> The issue here seems to be that vmwgfx always had some buffers that
> didn't immediately go through vmw_validation_bo_reserve. What's
> happening is that in vmwgfx_execbuf.c in vmw_execbuf_process we call
> vmw_validation_bo_reserve and after it we call
> vmw_validation_res_validate. Inside vmw_validation_res_validate (in
> vmwgfx_validation.c) we call vmw_resource_validate, which calls
> vmw_resource_do_validate . vmw_resource_do_validate has this code "ret
> = func->create(res);" which is an issue for vmwgfx_cotable.c . The
> func->create for cotable's is vmw_cotable_create which calls
> vmw_cotable_resize which creates, reserves and validates a new bo.

Just wow! No wonder that I missed that, I would have never ever been 
able to find it.

>
> In short a new bo is created in vmw_cotable_resize between
> ttm_eu_reserve_buffers and ttm_eu_fence_buffer_objects calls.

Well then the fix is trivial, we just need to call 
dma_resv_reserve_fences() on this new BO to make sure that at least one 
fence slot is reserved.

I will try to come up with a patch. Thanks a lot for that!

Regards,
Christian.

>
> z