On 19/12/2023 12:32 am, Joe Tretter wrote: > Hi Andrew. > > Yes, running the scrypt unit test twice simultaneously made the error > occur faster. > This time, this wasn't required because it failed straight away - the > boot parameter is present:
Is it always the same test which fails, or is it random? Looking at https://github.com/Tarsnap/scrypt it's only a trivial piece of userspace crypto. The fact that running multiple instances makes it fail more easily points towards some kind of register handling issue, but the fact that it repros only under Xen, and even with eager-fpu (which isn't the default on AMD, sadly), is weird. Looking at the scrypt source, it has alternative routines for the AESNI and SHANI instruction groups. However, because it's a Zen1, we don't have a useful way of filtering visible for PV dom0 userspace. First of all, can you get the exact CPU model and microcode version. `head /proc/cpuinfo` will be enough. But while you're at it, can you include `xl dmesg` too just in case there's something obvious showing up there too. Thanks, ~Andrew
