2016-09-29 15:19+0200, Paolo Bonzini:
> On 29/09/2016 13:23, Radim Krčmář wrote:
>> QEMU 2.7 allowed EIM even in configurations that were forbidden in the
>> last patch because they were not working, like old KVM or userspace
>> APIC. In order to keep backward compatibility, we again allow guests to
>> misbehave in non-obvious ways, and make it the default.
>>
>> Signed-off-by: Radim Krčmář <[email protected]>
>
> Ugh, I misremembered that VTD_ECAP_EIM was not set in 2.7. :( Perhaps
> it's better to drop this patch...
I think that adding this backward compatibility hack into code that is
supposed to be developed is not a good idea.
2016-09-29 12:01-0300, Eduardo Habkost:
> If you break compatibility and fix it in separate patches, you
> break bisectability (even for people that are bisecting bugs
> unrelated to EIM).
I'd keep it as a separate patch and let maintainers decide whether they
want to squish or drop it.
> (But I still don't understand if patch 6/7 really breaks
> anything, or not.)
Nothing useful.
It "breaks" three cases:
1) If user configured
-machine kernel_irqchip=off -device intel_iommu,intremap=on
QEMU 2.7 pc-q35-2.7 enabled (broken) EIM, but 2.8 wouldn't, leading
to a different machine.
(The same with new KVM and split irqchip.)
2) If user had old KVM and configured
-machine kernel_irqchip=split -device intel_iommu,intremap=on
QEMU 2.7 pc-q35-2.7 enabled (broken) EIM, but after offline migration
to 2.8, QEMU would refuse to start.
3) If user started a pc-q35-2.7 with QEMU 2.8 on a new KVM, then they
could use cluster x2APIC without a problem, but the guest wouldn't
work after offline migration to QEMU 2.7 (I'm not sure if this case
is supported).
Luckily, the intel-iommu device doesn't support live migration. :)
