Re: [Qemu-devel] [PATCH 05/23] hyperv: ensure VP index equal to QEMU cpu_index

[Roman Kagan]

On Thu, Jun 15, 2017 at 03:27:28PM +0200, Igor Mammedov wrote:
> On Thu, 15 Jun 2017 15:41:08 +0300
> Roman Kagan <address@hidden> wrote:
> > On Wed, Jun 14, 2017 at 03:00:27PM +0200, Igor Mammedov wrote:
> > > On Wed, 14 Jun 2017 13:26:44 +0200
> > > Paolo Bonzini <address@hidden> wrote:
> > >   
> > > > On 14/06/2017 13:25, Roman Kagan wrote:  
> > > > >> The problem with that is that it will break as soon as we create
> > > > >> VCPUs in a different order.  Unsolvable on hosts that don't allow
> > > > >> HV_X64_MSR_VP_INDEX to be set, however.    
> > > > > Right, thanks for putting together a detailed explanation.
> > > > > 
> > > > > This was a thinko back then, not to have HV_X64_MSR_VP_INDEX 
> > > > > maintained
> > > > > by QEMU.  I'm going to post a patch to KVM fixing that.
> > > > > 
> > > > > Meanwhile QEMU needs a way to maintain its notion of vp_index that is
> > > > >   1) in sync with kernel's notion
> > > > >   2) also with kernels that don't support setting the msr
> > > > >   3) persistent across migrations
> > > > > 
> > > > > cpu_index looked like a perfect candidate.
> > > > >     
> > > > 
> > > > What you want is the APIC id,  
> > >   
> > > > which _is_ cpu_index but may not be in the  
> > > depending on topology cpu_index won't be the same as APIC ID/vcpu_id
> > > /AMDs odd core count/.  
> > 
> > So vcpu_id can be sparse?
> yes
> 
> I suppose kernel even has cpu->apic_id map somewhere that you can reuse
> instead of custom map hv_context.vp_index[]

Right, but this is not the problem.  Whether apic_id matches the
expected vp_index semantics is.

> > > > future.  But the APIC id is also the KVM vcpu_id, so there's no need to
> > > > have VP_INDEX maintained by QEMU.  
> > > agreed it'd be better to reuse vcpu_id/apic id as interface between
> > > qemu/kvm/guest instead of adding additional cpu_index concept in ABI  
> > 
> > Having consulted the spec, I'm not so confident any more this is the
> > right move.
> > 
> > > 7.8.1 Virtual Processor Index
> > > 
> > > Virtual processors are identified by using an index (VP index). The
> > > maximum number of virtual processors per partition supported by the
> > > current implementation of the hypervisor can be obtained through CPUID
> > > leaf 0x40000005. A virtual processor index must be less than the
> > > maximum number of virtual processors per partition.
> > 
> > This seems to imply that VP index should be dense.  As if they use it
> > directly as an index into an array whose length is equal to the max
> > number of vcpus.  (BTW the value we report for it is currently hardcoded
> > to 0x40 which probably needs fixing, too.)
> the question here is where "maximum number of virtual processors" comes from

We provide it in a corresponding cpuid leaf.

> if it's possible to tell guest value explicitly,
> we can use pcms->apic_id_limit for it.

But, depending on the apic_id encoding scheme, this can be arbitrarily
bigger than the actual maximum number of vcpus.  Which can result in
guest confusion or inefficiency.

> > I'm starting to drift back to adding SET_MSRS support to
> > HV_X64_MSR_VP_INDEX in KVM to delegate the control to QEMU, and having
> > QEMU use cpu_index as its value...  :-/

Roman.