Re: [Qemu-devel] [PATCH RFC qom-cpu-next 1/6] cpu: Register VMStateDescription through CPUState

[Eduardo Habkost]

On Thu, Feb 07, 2013 at 07:02:32PM +0100, Andreas Färber wrote:
> Am 07.02.2013 18:46, schrieb Eduardo Habkost:
> > On Thu, Feb 07, 2013 at 06:20:52PM +0100, Andreas Färber wrote:
> >> Am 07.02.2013 17:40, schrieb Eduardo Habkost:
> >>> On Sat, Feb 02, 2013 at 04:04:11PM +0100, Andreas Färber wrote:
> >>>> In comparison to DeviceClass::vmsd, CPU VMState is split in two,
> >>>> "cpu_common" and "cpu", and uses cpu_index as instance_id instead of -1.
> >>>> Therefore add a CPU-specific CPUClass::vmsd field.
> >>>>
> >>>> Unlike the legacy CPUArchState registration, rather register CPUState.
> >>>>
> >>>> Signed-off-by: Juan Quintela <address@hidden>
> >>>> Signed-off-by: Andreas Färber <address@hidden>
> >>>> ---
> >>>>  exec.c            |   13 +++++++++++--
> >>>>  include/qom/cpu.h |    3 +++
> >>>>  2 Dateien geändert, 14 Zeilen hinzugefügt(+), 2 Zeilen entfernt(-)
> >>>>
> >>>> diff --git a/exec.c b/exec.c
> >>>> index b85508b..5eee174 100644
> >>>> --- a/exec.c
> >>>> +++ b/exec.c
> >>>> @@ -219,7 +219,7 @@ void cpu_exec_init_all(void)
> >>>>  #endif
> >>>>  }
> >>>>  
> >>>> -#if defined(CPU_SAVE_VERSION) && !defined(CONFIG_USER_ONLY)
> >>>> +#if !defined(CONFIG_USER_ONLY)
> >>>>  
> >>>>  static int cpu_common_post_load(void *opaque, int version_id)
> >>>>  {
> >>>> @@ -266,6 +266,9 @@ CPUState *qemu_get_cpu(int index)
> >>>>  void cpu_exec_init(CPUArchState *env)
> >>>>  {
> >>>>      CPUState *cpu = ENV_GET_CPU(env);
> >>>> +#if !defined(CONFIG_USER_ONLY) && !defined(CPU_SAVE_VERSION)
> >>>> +    CPUClass *cc = CPU_GET_CLASS(cpu);
> >>>> +#endif
> >>>>      CPUArchState **penv;
> >>>>      int cpu_index;
> >>>>  
> >>>> @@ -290,10 +293,16 @@ void cpu_exec_init(CPUArchState *env)
> >>>>  #if defined(CONFIG_USER_ONLY)
> >>>>      cpu_list_unlock();
> >>>>  #endif
> >>>> -#if defined(CPU_SAVE_VERSION) && !defined(CONFIG_USER_ONLY)
> >>>> +#if !defined(CONFIG_USER_ONLY)
> >>>>      vmstate_register(NULL, cpu_index, &vmstate_cpu_common, env);
> >>>> +#if defined(CPU_SAVE_VERSION)
> >>>>      register_savevm(NULL, "cpu", cpu_index, CPU_SAVE_VERSION,
> >>>>                      cpu_save, cpu_load, env);
> >>>
> >>> What about introducing stub register_savevm() function in libqemustub.a,
> >>> so we can eliminate the CONFIG_USER_ONLY ifdefs?
> >>>
> >>> (we already have vmstate_register()/vmstate_unregister() stubs).
> >>
> >> register_savevm() itself is not the issue, it's the VMStateDescription
> >> itself with all its external references that would be quite some (IMO
> >> useless) work to make available to *-user...
> > 
> > Are you talking about the VMStateDescription struct declaration, or
> > about actually setting the vmsd field?
> 
> I'm talking about, e.g., #include "machine.c", i.e. VMStateDescription
> vmstate_something = { ... }; something_else = &vmstate_something;.
> 
> This broke horribly.
> 
> > The struct declaration is available even if CONFIG_USER_ONLY is set. See
> > qdev.c. It doesn't have any #ifdefs around
> > vmstate_register()/vmstate_unregister() calls.
> > 
> > I don't suggest we set the field to non-NULL if CONFIG_USER_ONLY is set
> > (that would be difficult and unnecessary).
> 
> That's what I was saying, yes.
> 
> >>>> +#else
> >>>
> >>> Do we have any architecture that sets CPU_SAVE_VERSION _and_ sets
> >>> cc->vmsd at the same time? If not, we could put the "if (cc->vmsd)
> >>> vmstate_register()" part outside any #ifdef/#else block.
> >>
> >> They shouldn't. When this series and any follow-up by Juan himself were
> >> applied, there would be no more CPU_SAVE_VERSION and all CPUs would
> >> register a vmsd one way or another. Targets to be converted include ppc,
> >> arm and sparc.
> >>
> >> Together with the final RFC patch in this series, doing it inside an
> >> #else block avoids repeating the lax checks that have led alpha,
> >> openrisc and a few others to not register things correctly.
> > 
> > What exactly were the lax checks that have led them to not register
> > things correctly?
> 
> In short: Lack of patch 6.
> 
> !defined(CPU_SAVE_VERSION) but implementing cpu_save/load when they
> should've #define'd CPU_SAVE_VERSION.
> 
> In turn I want to assure that when !defined(CPU_SAVE_VERSION)
> implementing cpu_save/load leads to build error.
> 
> > Would my suggestion below cause the same problems?
> > 
> >> This is the
> >> part of the patch that I adopted from Juan. I don't insist though.
> > 
> > I am OK with "#ifdef CPU_SAVE_VERSION" #ifdef because it is for legacy
> > code (and should be temporary, right?), but I think we can easily drop
> > the #ifdefs around all the other lines. I mean, we can easily make the
> > code look like this:
> > 
> > void cpu_exec_init(CPUArchState *env)
> > {
> >     CPUState *cpu = ENV_GET_CPU(env);
> >     CPUClass *cc = CPU_GET_CLASS(cpu);
> >     [...]
> > 
> >     vmstate_register(NULL, cpu_index, &vmstate_cpu_common, env);
> 
> &vmstate_cpu_common will break for linux-user.

Oops. Then what about:

#if !defined(CONFIG_USER_ONLY)
static const VMStateDescription vmstate_cpu_common { ... };
#define cpu_common_vmsd (&vmstate_cpu_common)
#else
#define cpu_common_vmsd NULL
#endif
[...]
     vmstate_register(NULL, cpu_index, cpu_common_vmsd, env);
[...]

This pattern is similar to what I suggested for the code that
initializes cc->vmsd on the target-specific class_init functions. I
don't really love the way it looks, but I prefer #ifdefs on declarative
parts of the code than inside functions. I wonder if we could simplify
it further.


> 
> > #if defined(CPU_SAVE_VERSION)
> >     register_savevm(NULL, "cpu", cpu_index, CPU_SAVE_VERSION,
> >                      cpu_save, cpu_load, env);
> > #endif
> >     if (cc->vmsd) {
> >         vmstate_register(NULL, cpu_index, cc->vmsd, cpu);
> >     }
> > }
> > 
> > If we want to catch architectures that don't set CPU_SAVE_VERSION but
> > also don't register any vmsd (is this the bug you are trying to catch?),
> > we could do this:
> > 
> > #if defined(CPU_SAVE_VERSION)
> >     register_savevm(NULL, "cpu", cpu_index, CPU_SAVE_VERSION,
> >                      cpu_save, cpu_load, env);
> > #elif !defined(CONFIG_USER_ONLY)
> >     assert(cc->vmsd);
> 
> We would need to leave out the #elif and assert cc->vmsd == NULL, but
> then it would address my concern, yes.
> 
> !defined(CPU_SAVE_VERSION) => cc->vmsd != NULL is not guaranteed:

OK. I just thought that was the bug you were trying to catch, but it was
a different problem, as you explained above.

> For 1.4 some unmigratable CPUs register via dc->vmsd instead.
> 
> Regards,
> Andreas
> 
> > #endif
> >     if (cc->vmsd) {
> >         vmstate_register(NULL, cpu_index, cc->vmsd, cpu);
> >     }
> > 
> > 
> >>
> >> Regards,
> >> Andreas
> >>
> >>>> +    if (cc->vmsd != NULL) {
> >>>> +        vmstate_register(NULL, cpu_index, cc->vmsd, cpu);
> >>>> +    }
> >>>> +#endif
> >>>>  #endif
> >>>>  }
> >>>>  
> >>>> diff --git a/include/qom/cpu.h b/include/qom/cpu.h
> >>>> index 46f2247..b870752 100644
> >>>> --- a/include/qom/cpu.h
> >>>> +++ b/include/qom/cpu.h
> >>>> @@ -43,6 +43,7 @@ typedef struct CPUState CPUState;
> >>>>   * @class_by_name: Callback to map -cpu command line model name to an
> >>>>   * instantiatable CPU type.
> >>>>   * @reset: Callback to reset the #CPUState to its initial state.
> >>>> + * @vmsd: State description for migration.
> >>>>   *
> >>>>   * Represents a CPU family or model.
> >>>>   */
> >>>> @@ -54,6 +55,8 @@ typedef struct CPUClass {
> >>>>      ObjectClass *(*class_by_name)(const char *cpu_model);
> >>>>  
> >>>>      void (*reset)(CPUState *cpu);
> >>>> +
> >>>> +    const struct VMStateDescription *vmsd;
> >>>>  } CPUClass;
> >>>>  
> >>>>  struct KVMState;
> >>>> -- 
> >>>> 1.7.10.4
> >>
> >> -- 
> >> SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany
> >> GF: Jeff Hawn, Jennifer Guild, Felix Imendörffer; HRB 16746 AG Nürnberg
> 
> -- 
> SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany
> GF: Jeff Hawn, Jennifer Guild, Felix Imendörffer; HRB 16746 AG Nürnberg

-- 
Eduardo