Re: [PATCH v4 12/16] hw/i386: Use the apicid handlers from X86MachineState

[Eduardo Habkost]

On Tue, Feb 25, 2020 at 09:16:32AM +0100, Igor Mammedov wrote:
> On Mon, 24 Feb 2020 17:31:49 -0500
> Eduardo Habkost <address@hidden> wrote:
> 
> > On Mon, Feb 24, 2020 at 11:58:09AM -0600, Babu Moger wrote:
> > > 
> > > 
> > > On 2/24/20 11:19 AM, Igor Mammedov wrote:  
> > > > On Thu, 13 Feb 2020 12:17:46 -0600
> > > > Babu Moger <address@hidden> wrote:
> > > >   
> > > >> Check and Load the apicid handlers from X86CPUDefinition if available.
> > > >> Update the calling convention for the apicid handlers.  
> > > > 
> > > > Previous and this patch look too complicated for the task at the hand.
> > > > In particular, cpu_x86_init_apicid_fns() from previous patch adds 1 more
> > > > reference to Machine into i386/cpu.c (even though it's just a helper 
> > > > function)
> > > > and I think un-necessary hooks to X86CPUDefinition (it's not really 
> > > > CPU's
> > > > businesses to make up APIC-IDs).
> > > > 
> > > > I'd rather do opposite and get rid of the last explicit dependency to
> > > > ms->smp.cpus from cpu.c. But well, it's out of scope of this series,
> > > > so for this series I'd just try to avoid adding more Machine 
> > > > dependencies.
> > > > 
> > > > All 11/16 does is basically using hooks as a switch "I'm EPYC" to
> > > > set epyc specific encoding topo routines.
> > > > 
> > > > It could be accomplished by a simple Boolean flag like
> > > >  X86CPUDefinition::use_epyc_apic_id_encoding
> > > > 
> > > > and then cpu_x86_init_apicid_fns() could be replaced with trivial
> > > > helper like:
> > > > 
> > > >   x86_use_epyc_apic_id_encoding(char *cpu_type)
> > > >   {
> > > >       X86CPUClass *xcc = ... cpu_type ...
> > > >       return xcc->model->cpudef->use_epyc_apic_id_encoding
> > > >   }
> > > > 
> > > > then machine could override default[1] hooks using this helper
> > > > as the trigger
> > > >   x86_cpus_init()
> > > >   {
> > > >       // no need in dedicated function as it's the only instance it's 
> > > > going to be called ever
> > > >       if (x86_use_epyc_apic_id_encoding(ms->cpu_type)) {
> > > >             x86ms->apicid_from_cpu_idx = ...epyc...
> > > >             x86ms->topo_ids_from_apicid = ...epyc...
> > > >             x86ms->apicid_from_topo_ids = ...epyc...
> > > >             x86ms->apicid_pkg_offset = ...epyc...

[1]

> > > >       }
> > > >   }
> > > > 
> > > > That would be less invasive and won't create non necessary 
> > > > dependencies.  
> > > 
> > > Yes. We can achieve the task here with your approach mentioned above. But,
> > > we still will have a scaling issue. In future if a "new cpu model" comes
> > > up its own decoding, then we need to add another bolean flag use_new
> > > _cpu_apic_id_encoding. And then do that same check again. In that sense,
> > > the current approach is bit generic. Lets also hear from Eduardo.
> 
> Without another encoding on horizon, it looks like over-engineering.
> So lets think of a more generic approach later when need arises.
>  
> 
> > To be honest, I really hope the number of APIC ID initialization
> > variations won't grow in the future.
> > In either case, X86MachineState really doesn't seem to be the
> > right place to save the function pointers.  Whether we choose a
> > boolean flag or a collection of function pointers, model-specific
> > information belong to x86CPUClass and/or X86CPUDefinition, not
> > MachineState.
> 
> That's where I disagree, generating APIC ID and it's assignment
> shouldn't be part of CPU model. On real machines it's done by
> board/firmware, there board is purpose build for specific CPU.
> 
> The same should be in QEMU case, where universal QEMU board
> adapts its APIC initialization to used CPU and not other way
> around (i.e. it's not CPU's job to generate IDs).
> So hooks in machine look like a good approach to me.
> 
> I's fine to add small helper to CPU code to help board decide
> what APIC encoding to use, but I strongly disagree in putting
> more logic/data than that into CPU model.
> 
> What CPU does inside (I mean CPUID handling) it's separate
> business and in that case CPU usually knows what it's encoding
> and can use epyc/non_epyc functions directly if necessary.

I see your point.  I was assuming that the function pointers
would stay in cpu.c/X86CPUDefinition, then it wouldn't make sense
to copy the pointers to X86MachineClass.

If the functions are going to live in pc.c, then adding the
function pointers to X86MachineState might make sense and your
suggestion above[1] sounds reasonable.

-- 
Eduardo