On 11/29/21 12:39 PM, Woodhouse, David wrote:
> On Fri, 2021-07-23 at 13:29 +0200, Claudio Fontana wrote:
>> static void kvm_cpu_instance_init(CPUState *cs)
>> {
>> X86CPU *cpu = X86_CPU(cs);
>> + X86CPUClass *xcc = X86_CPU_GET_CLASS(cpu);
>>
>> host_cpu_instance_init(cpu);
>>
>> - if (!kvm_irqchip_in_kernel()) {
>> - x86_cpu_change_kvm_default("x2apic", "off");
>> - } else if (kvm_irqchip_is_split() && kvm_enable_x2apic()) {
>> - x86_cpu_change_kvm_default("kvm-msi-ext-dest-id", "on");
>> - }
>> -
>> - /* Special cases not set in the X86CPUDefinition structs: */
>> + if (xcc->model) {
>> + /* only applies to builtin_x86_defs cpus */
>> + if (!kvm_irqchip_in_kernel()) {
>> + x86_cpu_change_kvm_default("x2apic", "off");
>> + } else if (kvm_irqchip_is_split() && kvm_enable_x2apic()) {
>> + x86_cpu_change_kvm_default("kvm-msi-ext-dest-id", "on");
>> + }
>>
>> - x86_cpu_apply_props(cpu, kvm_default_props);
>> + /* Special cases not set in the X86CPUDefinition structs: */
>> + x86_cpu_apply_props(cpu, kvm_default_props);
>> + }
>>
>
> I think this causes a regression in x2apic and kvm-msi-ext-dest-id
> support. If you start qemu thus:
If I recall correctly, this change just tries to restore the behavior prior to
commit f5cc5a5c168674f84bf061cdb307c2d25fba5448 ,
fixing the issue introduced with the refactoring at that time.
Can you try bisecting prior to f5cc5a5c168674f84bf061cdb307c2d25fba5448 , to
see if the actual breakage comes from somewhere else?
>
> qemu-system-x86_64 -machine q35,accel=kvm,usb=off,kernel_irqchip=split -cpu
> host -smp 288,sockets=9,cores=16,threads=2
>
> The guest now sees those features, but we don't actually call
> kvm_enable_x2apic() so the APIC broadcast quirk doesn't get disabled,
> and interrupts targeted at APIC ID 255 are interpreted as broadcasts:
>
> [ 73.198504] __common_interrupt: 0.34 No irq handler for vector
> [ 73.198515] __common_interrupt: 11.34 No irq handler for vector
> [ 73.198517] __common_interrupt: 12.34 No irq handler for vector
> [ 73.198521] __common_interrupt: 15.34 No irq handler for vector
> [ 73.198524] __common_interrupt: 17.34 No irq handler for vector
> [ 73.198528] __common_interrupt: 34.34 No irq handler for vector
> [ 73.198529] __common_interrupt: 20.34 No irq handler for vector
> [ 73.198533] __common_interrupt: 41.34 No irq handler for vector
> [ 73.198539] __common_interrupt: 27.34 No irq handler for vector
> [ 73.198542] __common_interrupt: 28.34 No irq handler for vector
>
>
Any image to specifically test out? Would an actual 9 sockets machine be
required to reproduce this?
Thanks,
Claudio
