Re: [kvm-unit-tests PATCH v2 2/2] x86/run: Use separate "-accel" option to configure the accelerator

[Xiaoyao Li <xiaoyao.li@intel.com>]

On 9/30/2026 8:52 AM, Sean Christopherson wrote:
> On Tue, Sep 29, 2026, Xiaoyao Li wrote:
>> When configuring accelerator via machine's "accel" property, it cannot
>> pass additional ACCEL_PROPS. For example, it hits following error when
>> configuring KVM with "notify-vmexit":
>>
>>   # ACCEL="kvm,notify-vmexit=internal-error" ./x86-run x86/apic.c
>>   #   qemu-system-x86_64: Property 'pc-i440fx-11.1-machine.notify-vmexit' not found
>>
>> Switch to use separate "-accel" option to configure the accelerator.
> 
> FYI, this breaks the intel_iommu test when running with ACCEL="kvm,kernel-irqchip=on"
> because that's a split-only test:
> 
>   qemu_params = -M q35,kernel-irqchip=split -device intel-iommu,intremap=on,eim=off -device edu
> 
> and QEMU doesn't munge things together:
> 
>   qemu: -device intel-iommu,intremap=on,eim=off: Interrupt Remapping cannot work with kernel-irqchip=on, please use 'split|off'.

Because before this patch, the final QEMU command with
ACCEL="kvm,kernel-irqchip=on" is

   ... -machine accel=kvm,kernel-irqchip=on -kernel x86/intel-iommu.flat -smp 4
-M q35,kernel-irqchip=split -device intel-iommu,intremap=on,eim=off -device edu

ACCEL="kvm,kernel-irqchip=on" is made into the first "-machine
accel=kvm,kernel-irqchip=on" and qemu_params is put at the end, which has "-M
q35,kernel-irqchip=split". So the later "-M q35,kernel-irqchip=split" overrides
the earlier "kernel-irqchip=on" and it actually runs with kenrnel-irqchip=split.

But after this patch, the QEMU command becomes

   ... -accel kvm,kernel-irqchip=on -kernel x86/intel-iommu.flat -smp 4 -M
q35,kernel-irqchip=split -device intel-iommu,intremap=on,eim=off -device edu

though "-M q35,kernel-irqchip=split" is put behind "-accel
kvm,kernel-irqchip=on", QEMU applies "-accel" after "machine's kernel-irqchip".
So the final kernel-irqchip=on, which leads to the failure of Interrupt Remapping.

> Given that I'm pretty sure I'm literally the only person the world that runs KUT
> with ACCEL="kvm,kernel-irqchip=on", I'm going to apply this as-is.  I can filter
> out the failures very easily, and unlike the CET/LAM issue, this doesn't cause
> problems for ACCEL="kvm,kernel-irqchip=split", i.e. there's no loss in coverage.
> 
> If this impacts someone else, or someone really wants to figure out how to splice
> all of this together, I'm happy to revisit things.  For now, I just want to
> unblock the notify-vmexit stuff.
> 
>> Signed-off-by: Xiaoyao Li <xiaoyao.li@intel.com>
>> ---
>>  x86/run | 2 +-
>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/x86/run b/x86/run
>> index dd38f14c871f..a68d469dd1b8 100755
>> --- a/x86/run
>> +++ b/x86/run
>> @@ -43,7 +43,7 @@ fi
>>  pit="-global kvm-pit.lost_tick_policy=discard"
>>  
>>  command="${qemu} --no-reboot -nodefaults $pit $pc_testdev -display none -serial stdio $pci_testdev"
>> -command+=" -machine accel=$ACCEL$ACCEL_PROPS"
>> +command+=" -accel $ACCEL$ACCEL_PROPS"
>>  if [ "${CONFIG_EFI}" != y ]; then
>>  	command+=" -kernel"
>>  fi
>> -- 
>> 2.53.0
>>
>>