Re: [Qemu-ppc] [RFC for-2.12 0/8] spapr: Add optional capabilities

[Laurent Vivier]

On 11/12/2017 08:08, David Gibson wrote:
...
> The more specific motivation for this is that POWER9 has bugs in its
> hardware transactional memory (HTM) implementation making it unsafe to
> migrate POWER8 guests to POWER9 if they use HTM.

I have tested a migration from a POWER8 host to POWER9 host.

Without cap-htm=off, kernel crashes on destination with:

[   38.042689] NIP [c00000000005acec] tm_save_sprs+0x0/0x1c

[   38.042784] LR [c000000000019e20] __switch_to+0x140/0x470
       ...

If I start both destination and source with cap-htm=off (POWER9 QEMU
explicitly asks for it), it works well (except if I have stress tool
running in the guest, in this case the tool hangs, it seems we are
always losing some interrupts...).

And a note (this is not related to this series):

P9 -> P8 migration doesn't work very well...
Guest kernel has some traces, like:

[  119.471815] list_add corruption. prev->next should be next (
c0000000014db628), but was c00000000121f6c0. (prev=c000000001212768).

[  119.471930] ------------[ cut here ]------------

[  119.471969] WARNING: CPU: 0 PID: 355 at lib/list_debug.c:33
__list_add+0xe4/0x110

but the system is usable.

If we stress the guest (with iozone) and migrate, we have:

[  246.241859] kernel BUG at kernel/timer.c:1105!

[  246.241895] Oops: Exception in kernel mode, sig: 5 [#1]

...
[  248.246340] Kernel panic - not syncing: Fatal exception in interrupt


and QEMU exits.

Thanks,
Laurent