Re: [Qemu-ppc] [PATCH v2 4/7] ppc: open code cpu creation for machine types

[Igor Mammedov]

On Mon, 4 Jul 2016 13:54:55 +1000
David Gibson <address@hidden> wrote:

> On Sat, Jul 02, 2016 at 10:33:33AM +0200, Greg Kurz wrote:
> > On Sat, 2 Jul 2016 13:36:22 +0530
> > Bharata B Rao <address@hidden> wrote:
> > 
> > > On Sat, Jul 02, 2016 at 12:41:48AM +0200, Greg Kurz wrote:
> > > > If we want to generate cpu_dt_id in the machine code, this must
> > > > occur before the cpu gets realized. We must open code the cpu
> > > > creation to be able to do this.
> > > > 
> > > > This patch just does that. It borrows some lines from previous
> > > > work from Bharata to handle the feature parsing.
> > > > 
> > > > Signed-off-by: Greg Kurz <address@hidden>
> > > > ---
> > > >  hw/ppc/ppc.c |   39 ++++++++++++++++++++++++++++++++++++++-
> > > >  1 file changed, 38 insertions(+), 1 deletion(-)
> > > > 
> > > > diff --git a/hw/ppc/ppc.c b/hw/ppc/ppc.c
> > > > index dc3d214009c5..57f4ddd073d0 100644
> > > > --- a/hw/ppc/ppc.c
> > > > +++ b/hw/ppc/ppc.c
> > > > @@ -32,6 +32,7 @@
> > > >  #include "sysemu/cpus.h"
> > > >  #include "hw/timer/m48t59.h"
> > > >  #include "qemu/log.h"
> > > > +#include "qapi/error.h"
> > > >  #include "qemu/error-report.h"
> > > >  #include "hw/loader.h"
> > > >  #include "sysemu/kvm.h"
> > > > @@ -1353,5 +1354,41 @@ PowerPCCPU *ppc_get_vcpu_by_dt_id(int
> > > > cpu_dt_id)
> > > > 
> > > >  PowerPCCPU *ppc_cpu_init(const char *cpu_model)
> > > >  {
> > > > -    return POWERPC_CPU(cpu_generic_init(TYPE_POWERPC_CPU,
> > > > cpu_model));
> > > > +    PowerPCCPU *cpu;
> > > > +    CPUClass *cc;
> > > > +    ObjectClass *oc;
> > > > +    gchar **model_pieces;
> > > > +    Error *err = NULL;
> > > > +
> > > > +    model_pieces = g_strsplit(cpu_model, ",", 2);
> > > > +    if (!model_pieces[0]) {
> > > > +        error_report("Invalid/empty CPU model name");
> > > > +        return NULL;
> > > > +    }
> > > > +
> > > > +    oc = cpu_class_by_name(TYPE_POWERPC_CPU, model_pieces[0]);
> > > > +    if (oc == NULL) {
> > > > +        error_report("Unable to find CPU definition: %s",
> > > > model_pieces[0]);
> > > > +        return NULL;
> > > > +    }
> > > > +
> > > > +    cpu = POWERPC_CPU(object_new(object_class_get_name(oc)));
> > > > +
> > > > +    cc = CPU_CLASS(oc);
> > > > +    cc->parse_features(CPU(cpu), model_pieces[1], &err);  
> > > 
> > > Igor is working on a patchset to convert -cpu features into
> > > global properties. IIUC, after that patchset, it is not
> > > recommended to parse the -cpu features for every CPU but do it
> > > only once.
> > > 
> > 
> > cpu_generic_init() in the current code also does the parsing, and
> > as the title says, this patch is just about open coding the
> > creation. I don't want to change behavior yet.
> > 
> > But yes, I agree that we should only parse features once and I'll
> > be more than happy to fix this in a followup patch, based on Igor's
> > work.
> > 
> > In the meantime, maybe I can add a comment stating that the parsing
> > should go away ?
> 
> Right.  But the thing is by open coding here, you're making two copies
> that need to be fixed instead of one, which increases the chances of
> error.
this patch matches what has been done for x86 target as a pert of
decoupling *-user mode from machine emulation.

> It seems like it would be safer to change the generic code so there's
> a new generic function which doesn't do the realize which we can use
> on ppc (and other platforms when/if they need it).
We've had it in x86 until recently but I've replaced it
cpu_generic_init(), please don't go that route.

There is not much to generalize here so far it's basically following
code
   cpu = object_new(cpu-type)
   parse-features(cpu)

   set_properties(cpu) /* optional machine specific */

   cpu->realize()

once parse-features refactoring is merged, there won't be anything cpu
specific left as parse-features(cpu) could be moved to generic code
and only very machine specific set_properties would be left which are
not generalizable.

> 
> Doing the change just on ppc by making our own copy of
> cpu_generic_init() seems more like to lead to future mistakes.
I wouldn't touch cpu_init()/cpu_generic_init() and leave it only for
*-user users (as it's been done for x86).

> 
> > > That is what I attempted here in the context of supporting compat
> > > cpu type for pseries-2.7:
> > > 
> > > https://www.mail-archive.com/address@hidden/msg381660.html
> > > 
> > 
> > Yeah and this is where I borrowed some lines. :)
> > 
> > > Regards,
> > > Bharata.
> > > 
> > 
> > Cheers.
> > 
>