On 14.10.2014 15:44, Адонай Элохим wrote:

I'm tinkering around in radeon DRM code as a hobby so if you have a couple of minutes, could you explain, how GTT works for Radeon module? I've seen mentions of it here and there on the Internet, and tried to ping developers on IRC, but the overall picture is still unreachable for me.

GTT (or GART) is a mechanism for making physically non-contiguous pages of system RAM accessible to the GPU in a linear address range. This is achieved using a translation table (the TT in GTT) which maps the linear range addresses to physical (bus) addresses.

The linear address range is managed by TTM (same TT), the radeon driver just provides the radeon_ttm_backend_bind/unbind() hooks to it for binding the pages corresponding to a buffer object (BO) into the linear range and for unbinding them from it.

I also looked through the code in radeon_object.c, radeon_ttm.c and radeon_gem.c. In my mind, the GEM and the TTM do the same thing, so why is there a need for GEM and TTM coupled?..

The functionalities of TTM and GEM are overlapping, but not identical. The radeon driver uses memory management functionality of TTM and userspace interface functionality of GEM.