Re: [Qemu-devel] [RFC PATCH v2 3/5] spapr: Set stable_cpu_id for threads of CPU cores

[David Gibson]

On Fri, Jul 08, 2016 at 12:59:59PM +0200, Igor Mammedov wrote:
> On Fri, 8 Jul 2016 17:39:52 +1000
> David Gibson <address@hidden> wrote:
> 
> > On Fri, Jul 08, 2016 at 12:11:12PM +0530, Bharata B Rao wrote:
> > > On Fri, Jul 08, 2016 at 03:24:13PM +1000, David Gibson wrote:  
> > > > On Thu, Jul 07, 2016 at 08:20:23PM +0530, Bharata B Rao wrote:  
> > > > > Conditonally set stable_cpu_id for CPU threads that are created as 
> > > > > part
> > > > > of spapr CPU cores. The use of stable_cpu_id is enabled for 
> > > > > pseries-2.7
> > > > > onwards.
> > > > > 
> > > > > Signed-off-by: Bharata B Rao <address@hidden>
> > > > > ---
> > > > >  hw/ppc/spapr_cpu_core.c | 7 +++++++
> > > > >  1 file changed, 7 insertions(+)
> > > > > 
> > > > > diff --git a/hw/ppc/spapr_cpu_core.c b/hw/ppc/spapr_cpu_core.c
> > > > > index b104778..0ec3513 100644
> > > > > --- a/hw/ppc/spapr_cpu_core.c
> > > > > +++ b/hw/ppc/spapr_cpu_core.c
> > > > > @@ -293,8 +293,15 @@ static void spapr_cpu_core_realize(DeviceState 
> > > > > *dev, Error **errp)
> > > > >      for (i = 0; i < cc->nr_threads; i++) {
> > > > >          char id[32];
> > > > >          obj = sc->threads + i * size;
> > > > > +        CPUState *cs;
> > > > >  
> > > > >          object_initialize(obj, size, typename);
> > > > > +        cs = CPU(obj);
> > > > > +
> > > > > +        /* Use core_id (which is actually cpu_dt_id) as stable CPU 
> > > > > id */
> > > > > +        if (cs->has_stable_cpu_id) {
> > > > > +            cs->stable_cpu_id = cc->core_id + i;
> > > > > +        }  
> > > > 
> > > > Testing cs->has_stable_cpu_id here in machine type specific code seems
> > > > really weird.  It's the machine type that knows whether it has a
> > > > stable ID to give to the CPU or not, rather than the other way around.
> > > > 
> > > > Since we haven't yet had a release with cpu cores, I think the right
> > > > thing is for cpu_core to unconditionally set the stable ID (and set
> > > > has_stable_id to true).  
> > > 
> > > Right, we can set cpu_stable_id unconditionally here since this code path
> > > (core realize) will be taken only for pseries-2.7 onwards. has_stable_id
> > > will get set as part of the property we defined in SPAPR_COMPAT_2_7.  
> > 
> > Hrm, that's true.  But when you describe it like that it sounds like a
> > really non-obvious and fragile dependency between different components.
> that's how compat stuff is typically done for devices,
> CPUs shouldn't be an exception. 
> (consistency with other devices helps here in long run)
>  
> > > > The backup path that does thread-based cpu
> > > > init, can set has_stable_id to false (if that's not the default).  
> > > 
> > > Default is off, but turning it on for 2.7 will be inherited by 2.6
> > > and others below. Hence I have code to explicitly disable this prop
> > > for 2.6 and below via SPAPR_COMPAT_2_6.  
> > 
> > This is all seeming terribly awkward.
> Typically default is set the way so new machine type doesn't have
> to enable it explicitly.
> 
> However the way it's done here helps not to touch/check every user
> that uses cpu_index, limiting series impact only to code that
> asks for it, it look a lot safer to got this rout for now.
> 
> 
> >  Can we try investigating a
> > different approach:
> > 
> >     1. Rename cpu_index to cpu_id, but it's still used in the same
> >        places it's used.
> > 
> >     2. Remove assumptions that cpu_id values are contiguous or
> >        dense
> >     
> >     3. Machine type decides whether it wants to populate the cpu_id
> >        values explicitly, or leave it to generic code to calculate
> >        them as cpu_index is calculated now.
> > 
> >     4. Ideally, generic code enforces that the machine type populates
> >        either all or none of the cpu_id values.
> > 
> > Does that seem workable?
> Ideally we will get there some day (and may be get rid of cpu_index 
> altogether),
> but for now it seems too invasive with a lot of chances to introduce non 
> obvious
> regression.

Yes, that's a risk.  But I'm basically no longer convinced that it's
any higher than the risk of the same thing with the current approach.

> So I'd keep approach used in this series.

-- 
David Gibson                    | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
                                | _way_ _around_!
http://www.ozlabs.org/~dgibson

signature.asc
Description: PGP signature