2016-12-14 17:15+0100, David Hildenbrand:
>> kvm_for_each_vcpu(i, vcpu, kvm)
>> if (kvm_apic_present(vcpu))
>> - max_id = max(max_id, kvm_apic_id(vcpu->arch.apic));
>> + max_id = max(max_id, kvm_x2apic_id(vcpu->arch.apic));
>>
>> new = kvm_kvzalloc(sizeof(struct kvm_apic_map) +
>> sizeof(struct kvm_lapic *) * ((u64)max_id + 1));
>> @@ -179,16 +189,23 @@ static void recalculate_apic_map(struct kvm *kvm)
>> struct kvm_lapic *apic = vcpu->arch.apic;
>> struct kvm_lapic **cluster;
>> u16 mask;
>> - u32 ldr, aid;
>> + u32 ldr;
>> + u8 xapic_id;
>> + u32 x2apic_id;
>>
>> if (!kvm_apic_present(vcpu))
>> continue;
>>
>> - aid = kvm_apic_id(apic);
>
> think I'd even prefer here a simple
>
> aid = kvm_xapic_id(apic);
> if (apic_x2apic_mode(apic))
> aid = kvm_x2apic_id(apic);
>
> that would keep changes minimal and I don't really see any benefit in the
> code when splitting handling up.
It is neccesassary to write an entry for both IDs and I wanted to split
it before [4/4], because doing both changes at once seemed hard to
grasp.
Putting it here didn't work well either ... is a separate patch for the
hunk below better, or would you prefer to have it in [4/4]?
> Patch 4 then simply can fixup setting code
>
> if (aid <= new->max_apic_id && !new->phys_map[aid])
> new->phys_map[aid] = apic;
>
> (if I am not missing some important corner case here)
The trick is that we want to do the following even in xAPIC mode:
new->phys_map[kvm_x2apic_id(apic)] = apic;
This is the main idea of the hotplug hack -- to allow unique addressing
of processors that were reset in xAPIC mode. (And I add a disgusting
"x2apic_id > 0xff" condition in [4/4], because we still allow guests to
change xAPIC IDs, which wouldn't play nice with this.)
Hardware does a superset of this, because it only looks at lower 8 bits
of the desination ID when delivering to xAPIC.
When kvm_x2apic_id(apic) != kvm_xapic_id(apic), then the APIC is in
xAPIC mode so we definitely want to keep xAPIC working, hence
if (!apic_x2apic_mode(apic))
new->phys_map[kvm_xapic_id(apic)] = apic;
Two writes are necessary.
And there can already be another_apic "kvm_x2apic_id(another_apic) ==
kvm_xapic_id(apic)" so we prevent hotplug from breaking existing x2APIC
setups by doing "!new->phys_map[aid]" when setting xAPIC ID.
I hope we get a better solution in the future, but it would have to be
done at hardware (QEMU) level, because even firmware (seabios) doesn't
have standard ways to deal with this situation ...
>> - ldr = kvm_lapic_get_reg(apic, APIC_LDR);
>> + xapic_id = kvm_xapic_id(apic);
>> + x2apic_id = kvm_x2apic_id(apic);
>>
>> - if (aid <= new->max_apic_id)
>> - new->phys_map[aid] = apic;
>> + if (apic_x2apic_mode(apic) &&
>> + x2apic_id <= new->max_apic_id)
>> + new->phys_map[x2apic_id] = apic;
>> + else if (!apic_x2apic_mode(apic))
>
>
> This looks good to me.
