On Tue, Nov 15, 2022 at 09:22:53AM +0100, Arnd Bergmann wrote:
Have you tried what happens if you run the same thing on an x86 machine? I would expect them to behave the same way, but it's possible something goes wrong with the guest CPU if this ends up using some (but not all) of the logic from KVM that would use '-cpu host' instead of '-cpu max'. Note that the Neoverse CPU in the Altra machine does not support SVE.
I'm finding it hard to think of a failure pattern that would make it through VL discovery then fail at runtime but also not obviously trigger any issues in syscall-abi...
Other things you could easily try would use the same command line as above, with the possible combinations of '-cpu host' (replacing -cpu max) and '-enable-kvm'. Do you always get the same result?
The machine parameter accel={tcg,kvm} is useful for forcing a specific backend - it's probably wise to force TCG if you might be running on a job on a native architecture.
BTW there's some other funky stuff going on with that job, the syscall-abi test is stopped with a timeout after 45 seconds (as is sve-ptrace) which appears to be coming from a harness somewhere. The selection of FP tests run seems to miss fp-stress too.