Re: [PATCH v3] hw/i386/cpu: remove default_cpu_version and simplify

[Ani Sinha]

On Tue, Jan 21, 2025 at 4:58 PM Zhao Liu <zhao1.liu@intel.com> wrote:
>
> Hi Ani,
>
> Sorry for late reply.
>
> On Thu, Jan 16, 2025 at 09:04:18AM +0530, Ani Sinha wrote:
> > Date: Thu, 16 Jan 2025 09:04:18 +0530
> > From: Ani Sinha <anisinha@redhat.com>
> > Subject: [PATCH v3] hw/i386/cpu: remove default_cpu_version and simplify
> > X-Mailer: git-send-email 2.45.2
> >
> > commit 0788a56bd1ae3 ("i386: Make unversioned CPU models be aliases")
> > introduced 'default_cpu_version' for PCMachineClass. This created three
> > categories of CPU models:
> >  - Most unversioned CPU models would use version 1 by default.
> >  - For machines 4.0.1 and older that do not support cpu model aliases, a
> >    special default_cpu_version value of CPU_VERSION_LEGACY is used.
> >  - It was thought that future machines would use the latest value of cpu
> >    versions corresponding to default_cpu_version value of
> >    CPU_VERSION_LATEST [1].
> >
> > All pc machines still use the default cpu version of 1 for
> > unversioned cpu models. CPU_VERSION_LATEST is a moving target and
> > changes with time. Therefore, if machines use CPU_VERSION_LATEST, it would
> > mean that over a period of time, for the same machine type, the cpu version
> > would be different depending on what is latest at that time. This would
> > break guests even when they use a constant machine type. Therefore, for
> > pc machines, use of CPU_VERSION_LATEST is not possible. Currently, only
> > microvms use CPU_VERSION_LATEST.
> >
> > This change cleans up the complicated logic around default_cpu_version
> > including getting rid of default_cpu_version property itself. A couple of 
> > new
> > flags are introduced, one for the legacy model for machines 4.0.1 and older
> > and other for microvms. For older machines, a new pc machine property is
> > introduced that separates pc machine versions 4.0.1 and older from the newer
> > machines. 4.0.1 and older machines are scheduled to be deleted towards
> > end of 2025 since they would be 6 years old by then. At that time, we can
> > remove all logic around legacy cpus. Microvms are the only machines that
> > continue to use the latest cpu version. If this changes later, we can
> > remove all logic around x86_cpu_model_last_version(). Default cpu version
> > for unversioned cpu models is hardcoded to the value 1 and applies
> > unconditionally for all pc machine types of version 4.1 and above.
> >
> > This change also removes all complications around CPU_VERSION_AUTO
> > including removal of the value itself.
>
> I like the idea to remove CPU_VERSION_AUTO. Though this patch introduces
> 2 more new static variables ("use_legacy_cpu" and "use_last_cpu_version"),
> as you said, once 4.0.1 and older machines are removed, it's easy to
> clean up "use_legacy_cpu".
>
> > 1) See commit dcafd1ef0af227 ("i386: Register versioned CPU models")
> >
> > CC: imammedo@redhat.com
> > Signed-off-by: Ani Sinha <anisinha@redhat.com>
> > ---
>
> [snip]
>
> > -void x86_cpus_init(X86MachineState *x86ms, int default_cpu_version)
> > +void x86_legacy_cpus_init(X86MachineState *x86ms)
> > +{
> > +    machine_uses_legacy_cpu();
> > +    x86_cpus_init(x86ms);
> > +}
> > +
> > +void x86_cpus_init_with_latest_cpu_version(X86MachineState *x86ms)
> > +{
> > +    x86_cpu_uses_lastest_version();
> > +    x86_cpus_init(x86ms);
> > +}
>
> Could we simplify it even further, i.e., omit these two new helpers and
> just add x86_cpu_uses_lastest_version() and machine_uses_legacy_cpu() to
> the initialization of the PC & microvm, e.g.,
>
> --- a/hw/i386/microvm.c
> +++ b/hw/i386/microvm.c
> @@ -458,7 +458,8 @@ static void microvm_machine_state_init(MachineState 
> *machine)
>
>      microvm_memory_init(mms);
>
> -    x86_cpus_init_with_latest_cpu_version(x86ms);
> +    x86_cpu_uses_lastest_version();
> +    x86_cpus_init(x86ms);
>
>      microvm_devices_init(mms);
>  }
>
> and
>
> --- a/include/hw/i386/pc.h
> +++ b/include/hw/i386/pc.h
> @@ -138,11 +138,10 @@ static inline void pc_init_cpus(MachineState *ms)
>
>      if (pcmc->no_versioned_cpu_model) {
>          /* use legacy cpu as it does not support versions */
> -        x86_legacy_cpus_init(x86ms);
> -    } else {
> -        /* use non-legacy cpus */
> -        x86_cpus_init(x86ms);
> +        machine_uses_legacy_cpu();
>      }
> +
> +    x86_cpus_init(x86ms);
>  }

yeah this simplifies things a bit.

>
>  /* ioapic.c */
>
> [snip]
>
> > diff --git a/include/hw/i386/pc.h b/include/hw/i386/pc.h
> > index a558705cb9..ad43a233d8 100644
> > --- a/include/hw/i386/pc.h
> > +++ b/include/hw/i386/pc.h
> > @@ -92,9 +92,6 @@ struct PCMachineClass {
> >
> >      /* Compat options: */
> >
> > -    /* Default CPU model version.  See x86_cpu_set_default_version(). */
> > -    int default_cpu_version;
> > -
> >      /* ACPI compat: */
> >      bool has_acpi_build;
> >      int pci_root_uid;
> > @@ -125,11 +122,29 @@ struct PCMachineClass {
> >       * check for memory.
> >       */
> >      bool broken_32bit_mem_addr_check;
> > +
> > +    /* whether the machine supports versioned cpu models */
> > +    bool no_versioned_cpu_model;
> >  };
> >
> >  #define TYPE_PC_MACHINE "generic-pc-machine"
> >  OBJECT_DECLARE_TYPE(PCMachineState, PCMachineClass, PC_MACHINE)
> >
> > +static inline void pc_init_cpus(MachineState *ms)
>
> I think there's no need to declare as `inline`.

yes, otherwise we see failure like:

In file included from /workspace/qemu-ani/include/hw/xen/xen-x86.h:11,
                 from ../stubs/xen-hw-stub.c:11:
/workspace/qemu-ani/include/hw/i386/pc.h:133:13: error: ‘pc_init_cpus’
defined but not used [-Werror=unused-function]
  133 | static void pc_init_cpus(MachineState *ms)
      |             ^~~~~~~~~~~~

>
> > +{
> > +    X86MachineState *x86ms = X86_MACHINE(ms);
> > +    PCMachineState *pcms = PC_MACHINE(ms);
> > +    PCMachineClass *pcmc = PC_MACHINE_GET_CLASS(pcms);
> > +
> > +    if (pcmc->no_versioned_cpu_model) {
> > +        /* use legacy cpu as it does not support versions */
> > +        x86_legacy_cpus_init(x86ms);
> > +    } else {
> > +        /* use non-legacy cpus */
> > +        x86_cpus_init(x86ms);
> > +    }
> > +}
> > +
> >  /* ioapic.c */
>
> As my comment above, we can just call machine_uses_legacy_cpu()
> if pcmc->no_versioned_cpu_model is true.
>
> [snip]
>
> > -/*
> > - * We resolve CPU model aliases using -v1 when using "-machine
> > - * none", but this is just for compatibility while libvirt isn't
> > - * adapted to resolve CPU model versions before creating VMs.
> > - * See "Runnability guarantee of CPU models" at
> > - * docs/about/deprecated.rst.
> > - */
> > -X86CPUVersion default_cpu_version = 1;
> > +static bool use_legacy_cpu;
> > +void machine_uses_legacy_cpu(void)
>
> What about this name, "x86_cpu_set_legacy_version"?
>
> > +{
> > +    use_legacy_cpu = true;
> > +}
> >
> > -void x86_cpu_set_default_version(X86CPUVersion version)
> > +static bool use_last_cpu_version;
>
> Maybe "use_lastest_cpu"? Keep it in the same style as "use_legacy_cpu".
>
> > +void x86_cpu_uses_lastest_version(void)
>
> Similarly, What about "x86_cpu_set_latest_version"?
>
> >  {
> > -    /* Translating CPU_VERSION_AUTO to CPU_VERSION_AUTO doesn't make sense 
> > */
> > -    assert(version != CPU_VERSION_AUTO);
> > -    default_cpu_version = version;
> > +    use_last_cpu_version = true;
> >  }
> >
> >  static X86CPUVersion x86_cpu_model_last_version(const X86CPUModel *model)
> > @@ -5376,14 +5375,11 @@ static X86CPUVersion 
> > x86_cpu_model_last_version(const X86CPUModel *model)
> >  /* Return the actual version being used for a specific CPU model */
> >  static X86CPUVersion x86_cpu_model_resolve_version(const X86CPUModel 
> > *model)
> >  {
> > -    X86CPUVersion v = model->version;
> > -    if (v == CPU_VERSION_AUTO) {
> > -        v = default_cpu_version;
> > -    }
> > -    if (v == CPU_VERSION_LATEST) {
> > +    if (use_last_cpu_version) {
> >          return x86_cpu_model_last_version(model);
> >      }
> > -    return v;
> > +
> > +    return model->version;
> >  }
> >
> >  static const Property max_x86_cpu_properties[] = {
> > @@ -5987,6 +5983,12 @@ static char *x86_cpu_class_get_alias_of(X86CPUClass 
> > *cc)
> >      if (!cc->model || !cc->model->is_alias) {
> >          return NULL;
> >      }
> > +
> > +    if (use_legacy_cpu) {
> > +        /* legacy cpu models do not support cpu aliases */
> > +        return NULL;
> > +    }
> > +
> >      version = x86_cpu_model_resolve_version(cc->model);
> >      if (version <= 0) {
>
> I understand this non-NULL check is origianl for legacy CPU version.

yes good catch,

> So it's also necessary to remove it, or convert it to "assert(version)"?

good idea.

>
> >          return NULL;
> > @@ -6004,11 +6006,7 @@ static void x86_cpu_list_entry(gpointer data, 
> > gpointer user_data)
> >      g_autofree char *model_id = x86_cpu_class_get_model_id(cc);
> >
> >      if (!desc && alias_of) {
> > -        if (cc->model && cc->model->version == CPU_VERSION_AUTO) {
> > -            desc = g_strdup("(alias configured by machine type)");
> > -        } else {
> >              desc = g_strdup_printf("(alias of %s)", alias_of);
> > -        }
> >      }
> >      if (!desc && cc->model && cc->model->note) {
> >          desc = g_strdup_printf("%s [%s]", model_id, cc->model->note);
> > @@ -6115,7 +6113,7 @@ static void x86_cpu_definition_entry(gpointer data, 
> > gpointer user_data)
> >       * Old machine types won't report aliases, so that alias translation
> >       * doesn't break compatibility with previous QEMU versions.
> >       */
> > -    if (default_cpu_version != CPU_VERSION_LEGACY) {
> > +    if (!use_legacy_cpu) {
> >          info->alias_of = x86_cpu_class_get_alias_of(cc);
> >      }
>
> Do we need the check of "!use_legacy_cpu"?
>
> x86_cpu_class_get_alias_of() returns NULL if use_legacy_cpu is true.

another good catch.

I have sent v4.