Re: [Qemu-devel] [RFC PATCH v1 05/10] cpu: Abstract CPU core type

[David Gibson]

On Wed, Mar 09, 2016 at 11:32:28AM +0100, Igor Mammedov wrote:
> On Wed, 9 Mar 2016 13:55:51 +1100
> David Gibson <address@hidden> wrote:
> 
> > On Tue, Mar 08, 2016 at 10:11:17AM +0100, Igor Mammedov wrote:
> > > On Tue, 8 Mar 2016 14:57:10 +1100
> > > David Gibson <address@hidden> wrote:
> > >   
> > > > On Mon, Mar 07, 2016 at 11:40:11AM +0100, Igor Mammedov wrote:  
> > > > > On Mon, 7 Mar 2016 14:01:55 +0530
> > > > > Bharata B Rao <address@hidden> wrote:
> > > > >     
> > > > > > On Mon, Mar 07, 2016 at 02:36:55PM +1100, David Gibson wrote:    
> > > > > > > On Fri, Mar 04, 2016 at 07:07:20PM +0100, Igor Mammedov wrote:    
> > > > > > >   
> > > > > > > > On Fri, 4 Mar 2016 16:32:53 +0530
> > > > > > > > Bharata B Rao <address@hidden> wrote:
> > > > > > > >       
> > > > > > > > > On Fri, Mar 04, 2016 at 11:38:45AM +0100, Igor Mammedov 
> > > > > > > > > wrote:      
> > > > > > > > > > On Fri,  4 Mar 2016 12:24:16 +0530
> > > > > > > > > > Bharata B Rao <address@hidden> wrote:
> > > > > > > > > >         
> > > > > > > > > > > Add an abstract CPU core type that could be used by 
> > > > > > > > > > > machines that want
> > > > > > > > > > > to define and hotplug CPUs in core granularity.
> > > > > > > > > > > 
> > > > > > > > > > > Signed-off-by: Bharata B Rao <address@hidden>
> > > > > > > > > > > ---
> > > > > > > > > > >  hw/cpu/Makefile.objs  |  1 +
> > > > > > > > > > >  hw/cpu/core.c         | 44 
> > > > > > > > > > > ++++++++++++++++++++++++++++++++++++++++++++
> > > > > > > > > > >  include/hw/cpu/core.h | 30 ++++++++++++++++++++++++++++++
> > > > > > > > > > >  3 files changed, 75 insertions(+)
> > > > > > > > > > >  create mode 100644 hw/cpu/core.c
> > > > > > > > > > >  create mode 100644 include/hw/cpu/core.h
> > > > > > > > > > > 
> > > > > > > > > > > diff --git a/hw/cpu/Makefile.objs b/hw/cpu/Makefile.objs
> > > > > > > > > > > index 0954a18..942a4bb 100644
> > > > > > > > > > > --- a/hw/cpu/Makefile.objs
> > > > > > > > > > > +++ b/hw/cpu/Makefile.objs
> > > > > > > > > > > @@ -2,4 +2,5 @@ obj-$(CONFIG_ARM11MPCORE) += arm11mpcore.o
> > > > > > > > > > >  obj-$(CONFIG_REALVIEW) += realview_mpcore.o
> > > > > > > > > > >  obj-$(CONFIG_A9MPCORE) += a9mpcore.o
> > > > > > > > > > >  obj-$(CONFIG_A15MPCORE) += a15mpcore.o
> > > > > > > > > > > +obj-y += core.o
> > > > > > > > > > >  
> > > > > > > > > > > diff --git a/hw/cpu/core.c b/hw/cpu/core.c
> > > > > > > > > > > new file mode 100644
> > > > > > > > > > > index 0000000..d8caf37
> > > > > > > > > > > --- /dev/null
> > > > > > > > > > > +++ b/hw/cpu/core.c
> > > > > > > > > > > @@ -0,0 +1,44 @@
> > > > > > > > > > > +/*
> > > > > > > > > > > + * CPU core abstract device
> > > > > > > > > > > + *
> > > > > > > > > > > + * Copyright (C) 2016 Bharata B Rao <address@hidden>
> > > > > > > > > > > + *
> > > > > > > > > > > + * This work is licensed under the terms of the GNU GPL, 
> > > > > > > > > > > version 2 or later.
> > > > > > > > > > > + * See the COPYING file in the top-level directory.
> > > > > > > > > > > + */
> > > > > > > > > > > +#include "hw/cpu/core.h"
> > > > > > > > > > > +
> > > > > > > > > > > +static char *core_prop_get_slot(Object *obj, Error 
> > > > > > > > > > > **errp)
> > > > > > > > > > > +{
> > > > > > > > > > > +    CPUCore *core = CPU_CORE(obj);
> > > > > > > > > > > +
> > > > > > > > > > > +    return g_strdup(core->slot);
> > > > > > > > > > > +}
> > > > > > > > > > > +
> > > > > > > > > > > +static void core_prop_set_slot(Object *obj, const char 
> > > > > > > > > > > *val, Error **errp)
> > > > > > > > > > > +{
> > > > > > > > > > > +    CPUCore *core = CPU_CORE(obj);
> > > > > > > > > > > +
> > > > > > > > > > > +    core->slot = g_strdup(val);
> > > > > > > > > > > +}
> > > > > > > > > > > +
> > > > > > > > > > > +static void cpu_core_instance_init(Object *obj)
> > > > > > > > > > > +{
> > > > > > > > > > > +    object_property_add_str(obj, "slot", 
> > > > > > > > > > > core_prop_get_slot, core_prop_set_slot,
> > > > > > > > > > > +                            NULL);
> > > > > > > > > > > +}
> > > > > > > > > > > +
> > > > > > > > > > > +static const TypeInfo cpu_core_type_info = {
> > > > > > > > > > > +    .name = TYPE_CPU_CORE,
> > > > > > > > > > > +    .parent = TYPE_DEVICE,
> > > > > > > > > > > +    .abstract = true,
> > > > > > > > > > > +    .instance_size = sizeof(CPUCore),
> > > > > > > > > > > +    .instance_init = cpu_core_instance_init,
> > > > > > > > > > > +};
> > > > > > > > > > > +
> > > > > > > > > > > +static void cpu_core_register_types(void)
> > > > > > > > > > > +{
> > > > > > > > > > > +    type_register_static(&cpu_core_type_info);
> > > > > > > > > > > +}
> > > > > > > > > > > +
> > > > > > > > > > > +type_init(cpu_core_register_types)
> > > > > > > > > > > diff --git a/include/hw/cpu/core.h b/include/hw/cpu/core.h
> > > > > > > > > > > new file mode 100644
> > > > > > > > > > > index 0000000..2daa724
> > > > > > > > > > > --- /dev/null
> > > > > > > > > > > +++ b/include/hw/cpu/core.h
> > > > > > > > > > > @@ -0,0 +1,30 @@
> > > > > > > > > > > +/*
> > > > > > > > > > > + * CPU core abstract device
> > > > > > > > > > > + *
> > > > > > > > > > > + * Copyright (C) 2016 Bharata B Rao <address@hidden>
> > > > > > > > > > > + *
> > > > > > > > > > > + * This work is licensed under the terms of the GNU GPL, 
> > > > > > > > > > > version 2 or later.
> > > > > > > > > > > + * See the COPYING file in the top-level directory.
> > > > > > > > > > > + */
> > > > > > > > > > > +#ifndef HW_CPU_CORE_H
> > > > > > > > > > > +#define HW_CPU_CORE_H
> > > > > > > > > > > +
> > > > > > > > > > > +#include "qemu/osdep.h"
> > > > > > > > > > > +#include "hw/qdev.h"
> > > > > > > > > > > +
> > > > > > > > > > > +#define TYPE_CPU_CORE "cpu-core"
> > > > > > > > > > > +
> > > > > > > > > > > +#define CPU_CORE(obj) \
> > > > > > > > > > > +    OBJECT_CHECK(CPUCore, (obj), TYPE_CPU_CORE)
> > > > > > > > > > > +
> > > > > > > > > > > +typedef struct CPUCore {
> > > > > > > > > > > +    /*< private >*/
> > > > > > > > > > > +    DeviceState parent_obj;
> > > > > > > > > > > +
> > > > > > > > > > > +    /*< public >*/
> > > > > > > > > > > +    char *slot;
> > > > > > > > > > > +} CPUCore;
> > > > > > > > > > > +
> > > > > > > > > > > +#define CPU_CORE_SLOT_PROP "slot"        
> > > > > > > > > > as it's generic property I'd rename to 'core' so it would 
> > > > > > > > > > fit all users        
> > > > > > > > > 
> > > > > > > > > Ok. Also note that this is a string property which is 
> > > > > > > > > associated with the
> > > > > > > > > link name (string) that we created from machine object to 
> > > > > > > > > this core. I think
> > > > > > > > > it would be ideal if this becomes an interger  property in 
> > > > > > > > > which case it
> > > > > > > > > becomes easier to feed the core location into your 
> > > > > > > > > CPUSlotProperties.core.      
> > > > > > > > agreed, it should be core number.      
> > > > > > > 
> > > > > > > The slot stuff is continuing to confuse me a bit.  I see that we 
> > > > > > > need
> > > > > > > some kind of "address" value, but how best to do it is not clear 
> > > > > > > to
> > > > > > > me.
> > > > > > > 
> > > > > > > Changing this to an integer sounds like it's probably a good idea.
> > > > > > > I'm a bit wary of just calling it "core" though.  Do all platforms
> > > > > > > even necessarily have a core id?
> > > > > > > 
> > > > > > > I'm wondering if the addressing is something that needs to move 
> > > > > > > the
> > > > > > > the platform specific subtypes, while some other stuff can move 
> > > > > > > to the
> > > > > > > generic base type.
> > > > > > >       
> > > > > > > > > > on top of that I'd add numeric 'threads' property to base 
> > > > > > > > > > class so
> > > > > > > > > > all derived cores would inherit it.
> > > > > > > > > > 
> > > > > > > > > > Then as easy integration with -smp threads=x, a machine 
> > > > > > > > > > could push
> > > > > > > > > > a global variable 'cpu-core.threads=[smp_threads]' which 
> > > > > > > > > > would
> > > > > > > > > > make every created cpu-core object to have threads set
> > > > > > > > > > at instance_init() time (device_init).
> > > > > > > > > > 
> > > > > > > > > > That way user won't have to specify 'threads=y' for every
> > > > > > > > > >   device_add spapr-core,core=x
> > > > > > > > > > as it will be taken from global property 'cpu-core.threads'
> > > > > > > > > > but if user wishes he/she still could override global by 
> > > > > > > > > > explicitly
> > > > > > > > > > providing thread property at device_add time:
> > > > > > > > > >   device_add spapr-core,core=x,threads=y
> > > > > > > > > > 
> > > > > > > > > > wrt this series it would mean, instead of creating threads 
> > > > > > > > > > in property
> > > > > > > > > > setter, delaying threads creation to core.realize() time,
> > > > > > > > > > but since realize is allowed to fail it should be fine do 
> > > > > > > > > > so.        
> > > > > > > > > 
> > > > > > > > > Ok that would suit us as there are two properties on which 
> > > > > > > > > thread creation
> > > > > > > > > is dependent upon: nr_threads and cpu_model. If thread 
> > > > > > > > > objects can be
> > > > > > > > > created at core realize time, then we don't have to resort to 
> > > > > > > > > the ugliness
> > > > > > > > > of creating the threads from either of the property setters. 
> > > > > > > > > I always
> > > > > > > > > assumed that we shouldn't be creating objects from realize, 
> > > > > > > > > but if that
> > > > > > > > > is fine, it is good.      
> > > > > > > > since realize is allowed to fail, it should be safe from 
> > > > > > > > hotplug pov
> > > > > > > > to create internal objects there, as far as proper cleanups are 
> > > > > > > > done
> > > > > > > > for failure path.      
> > > > > > > 
> > > > > > > Right, moving the "nr_threads" property to the base type seems 
> > > > > > > like a
> > > > > > > good idea to me.      
> > > > > > 
> > > > > > And we will also move the cpu_model property (now being tracked by
> > > > > > an ObjectClass pointer) to the base type ?    
> > > > > I'm not sure that moving cpu_model to the base class is the right 
> > > > > thing,
> > > > > I'd keep it local to platform for now.    
> > > > 
> > > > I tend to agree, although I'm not sure that I could really explain why
> > > > :/
> > > >   
> > > > > Could you have several spapr core types? One per CPU model?
> > > > > That way you won't need to track cpu_model when using device_add.    
> > > > 
> > > > We could in theory, but it would be pretty inconvenient.  Because this
> > > > is a paravirt platform, there really can't be any core-level
> > > > difference between them, and it would mean creating a fair batch of
> > > > core types for the various minor POWER7 and POWER8 variants - and
> > > > needing to update this whenever IBM makes a new version.  I suspect it
> > > > would also introduce more wrinkles in order to have a correct
> > > > "spapr-core-host" type matching the "HOST" cpu thread type.  Since KVM
> > > > (HV) only supports the HOST thread type, that's a fairly big issue.  
> > > Welcome to x86 world, that's roughly what we have there.  
> > 
> > I don't really follow you.  x86 doesn't have core devices at all for
> > the moment.
> > 
> > What I'm saying here is that using different core subtypes for every
> > cpu subtype on power would mean 2 types for each minor variant, rather
> > than just 1.
> I think the same applies to cpu_model and transitioning CPUs to -device.
> X86 also has a bunch of cpu_model-s which have some minor variations
> but it still generates a type per cpu_model.
> If some day we are to implement socket objects for x86 that would also
> mean to have a socket types per each cpu model.

Hmm.  I guess.

What concerns me is the possible combinatorial explosion of # cpu
models * # of machine types leading to an enormous number of
core/socket types.

The other thing is that the platform specific core types belong with
the machine type code, which means they can't naturally use macro
magic or whatever to generate them in parallel with the thread device
types in the core target-ppc code.

> As I understand cpu_model is a legacy option which should translate to
> a corresponding QOM type (CPU device) which could be used with
> -device/device_add.

Sure, but the "cpu_model" parameter could be "thread type name" just
as easily.

> As analogy, QEMU has legacy -net model=foo[1234] option, but when network 
> cards
> were converted to -device interface that in the end became a set of QOM types
> like -device foo[1234], it was easier in case of network cards as
> they where a separate devices models to begin with, the thing to note
> here is that they weren't converted to a single 'network_card' type with
> 'model' property.

-- 
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