Quoting Jonathan Cavitt (2023-02-17 17:20:19) > The gt_tlb live selftest has the same code coverage as the > igt_cs_tlb subtest of gtt. True, the intent of the code is the same, but gt_tlb has had a much high success rate at hitting TLB bugs, so is my preferred test. > However, igt_cs_tlb is hitting > an issue in that we are updating PTE in gt0->vm but > executing on the media tile without updating gt1->vm. I'm no longer convinced this a good explanation of the issue, as unlike the i915_requests selftest this is using a GEM context and not the local kernel contexts. The GEM context->vm should work equally on the mtl render and media tiles, afaict. The failure is very early in running on media tile (after running on the render tile) so I think it should be easy enough to reproduce in a simpler test to narrow down the cause. > This issue is corrected in gt_tlb, and thus igt_cs_tlb > is obsolete and should be removed. gt_tlb supersedes igt_cs_tlb, that I can agree on, Acked-by: Chris Wilson <chris.p.wilson@xxxxxxxxxxxxxxx> -Chris
