BE KVM breakage in kernel 5.14 and forward

BE KVM breakage in kernel 5.14 and forward

[Marcus Comstedt]

Greetings!

I'm using a Talos II with dual POWER9 DD2.2 processors in a BE, radix,
64k page config.

When trying to upgrade from 5.4.206 to 6.1.26, I noticed that all my
KVM virtual machines stopped working.  I tried the other stables, and
found that 5.15.111 was also broken, but 5.10.179 was fine.

So I did a bisect between 5.10 and 5.15 and ended up with the
following commit:


  commit 89d35b23910158a9add33a206e973f4227906d3c (HEAD)
  Author: Nicholas Piggin <npiggin@gmail.com>
  Date:   Fri May 28 19:07:34 2021 +1000

      KVM: PPC: Book3S HV P9: Implement the rest of the P9 path in C


In the commit before (9dc2babc185e) everything is working, but from
89d35b2391015 and forward what happens is that any KVM guest will
freeze really early (after the "FW Version = git-6b6c16b4b4076350"
printout from SLOF) running at 100% CPU on a single thread without
progress.  There is no relevant message in dmesg.

I notice that the commit in question seems to make some assumptions
related to the stack frame layout.  BE kernels are always
compiled with -mabi=elfv1 (hardcoded in arch/powerpc/Makefile for
5.14/5.15, then moved to arch/powerpc/platforms/Kconfig.cputype in
6.1); could it be that this code needs to be adjusted depending on the
ELF ABI version?



  // Marcus