Hi! On Mon, 2024-03-11 at 14:07 +0100, Thomas Hellström wrote: > On Fri, 2024-03-08 at 13:13 +0530, Somalapuram, Amaranath wrote: > > Patches are tested on AMD platform. > > Repeated stress test on Unigine Heaven, memory full (VRAM + GTT + > > system > > SWAP), then free. > > No errors/warning in kernel log. > > Any suggestion specific tests? > > We are testing locally against Intel Xe CI and Intel i915 CI which > should give rather good coverage. If there are some amdgpu tests that > exercise eviction / swapping also with a lot of local objects (Vulkan > apps?) that would be great. > > Thanks, > Thomas > Any updates on this? FWIW, For patch 3, IMO after looking a bit at other solutions, IMO this is the preferred solution mostly because it is self-contained. In particular if we allow drivers to iterate over the LRU lists with this interface, most likely if we add semantics like "You must block any bulk lru bumping if unlocking the lru_lock" That becomes pretty nasty and will most likely end up incorrect. It might well be that we've traversed well into a bulk move lru sublist before we try to unlock. /Thomas
