On Thursday, July 16, 2026 4:00:50 PM Central European Summer Time Alex > > > > Right, AFAIU assuming the CP hangs, instead of waiting for the TDR > > timeout with this patch it will be instant via drm_sched_fault() called > > from either irq handler. Or if CP does not hang but somehow ends up > > going over it with corruption with this patch it will insta GPU hang. > > > > In any case I think it is needed to have a definitive answer whether the > > interrupts were not enabled until now, and if so, also exercise the > > until now unused code paths just to make sure it all works as expected. > > Presumably there are test cases which trigger both conditions? > > > > Then I would suggest putting this info in the commit message. That is, > > instead of saying "These were used without ever calling get()/put() on > > them." expand with the full story so the whole situation is clear both > > before and after the patch. > > > > Assuming interrupts were not enabled until now, that the unused irq > > > handlers work as expected, and with the improved commit message: > The interrupts would not have been enabled if get() was never called. > I think they were never enabled because radeon didn't implement > support and this code was ported from radeon. > > > > Unless you don't have the hardware to test it and there are no existing > > test cases to verify it? > > An example test would be attempting to access a privileged register > from a user IB.
Unfortunately I don't have any test cases like that. We've had some plans for writing such test cases and integrating them in IGT but the person who was going to write them left and since then I haven't managed to get around to it yet. So I've mainly just used the hang test suite for testing that soft reset works. If you have doubts about these interrupts on GFX6 then I think we should just drop this patch until we finish those test cases. I still plan to write those test cases but it's really tedious and thankless work, so it hasn't been my priority thus far. Best regards, Timur
