Re: [PATCH] drm/radeon: Only flush HDP cache from idle ioctl if BO is in VRAM

On 05.08.2014 07:01, Marek Olšák wrote:

I'm afraid this won't always work and it can be a source of bugs.

Userspace doesn't have to call GEM_WAIT_IDLE before a CPU access to a VRAM buffer. For example, consider a wait-idle request with a non-zero timeout, which is implemented as a loop which calls GEM_BUSY. Also, userspace can use fences (alright they are backed by 1-page-sized VRAM buffers at the moment) and it may use real fences in the future which are not tied to a buffer object.

If the HDP flush isn't allowed in userspace IBs, I think we will have to expose it as an ioctl and call it explicitly from userspace.

I understand your concerns, but my patch doesn't change anything wrt them, does it?