On 3/8/21 3:57 PM, Philippe Mathieu-Daudé wrote: > On 3/8/21 2:52 PM, Claudio Fontana wrote: >> On 3/8/21 2:27 PM, Philippe Mathieu-Daudé wrote: >>> Hi Claudio, >>> >>> On 3/8/21 1:57 PM, Claudio Fontana wrote: >>>> Hi, >>>> >>>> anything else for me to do here? >>>> >>>> The latest rebased state of this series should be always available here: >>>> >>>> https://gitlab.com/hw-claudio/qemu/-/tree/i386_cleanup_8 >>>> >>>> When it comes to the ARM cleanup series, >>>> I would like to have the tests pass for ARM, before doing even more >>>> changes, could you help me there Philippe? >>>> >>>> Maybe applying some of your changes on top would fix the failures? I >>>> tried, for example with the arm-cpu-features ones, but it didn't work for >>>> me.. >>> >>> TBH I wrote these patches during my personal spare time and this >>> became a real Pandora box that drained too much energy. I prefer >>> to step back and focus on finishing smaller tasks before burning >>> out. That said I appreciate your effort and am interested in >>> following / reviewing your work. >>> >>> Regards, >>> >>> Phil. >>> >> >> Thanks Philippe for sharing this, and I agree completely, it is very >> draining. >> >> The effort of making tests happy that run in artificial environments in >> particular often feels to me >> as too disconnected from actually ensuring that there is no real run time >> regression. >> >> qtest_enabled() (implicitly, or explicitly via open-ended else statements) >> is another painful variable to keep in mind in cpu and machine code, so it >> is not helpful in my view. >> >> I'll try to push more to get the tests running again, if you have any >> comment or idea, feel free to just point me in the right direction, >> that is very valuable to me, even without working code. > > Basically I gave up after realizing from Daniel reviews that we need > QMP commands to query QEMU at runtime its built-in features, so we > have build-agnostic tests easier to maintain. I agree this is the > best way to resolve this particular case, but also scale for various > other cases. >
Well, yes, but in order to get things to work, even just a kvm-build fix would be good until we have the perfect solution, no? We also fixed the tcg tests when doing this for i386, so I think we can fix these issues for arm too. But this doesn't mean that we don't need the QMP commands to query QEMU at runtime for its "built-in"/module-loaded features. We need that too, and I suspect this will be more and more needed by libvirt, as QEMU modularizes. I just think the two things could proceed in parallel.. Ciao, Claudio
