Re: [Qemu-devel] [PATCH v2 02/10] machine: introduce MachineClass.possible_cpu_arch_ids() hook

[Igor Mammedov]

On Fri, 12 Feb 2016 16:33:28 -0200
Eduardo Habkost <address@hidden> wrote:

> On Fri, Feb 05, 2016 at 07:06:58PM +0100, Igor Mammedov wrote:
> > on x86 currently range 0..max_cpus is used to generate
> > architecture-dependent CPU ID (APIC Id) for each present
> > and possible CPUs. However architecture-dependent CPU IDs
> > list could be sparse and code that needs to enumerate
> > all IDs (ACPI) ended up doing guess work enumerating all
> > possible and impossible IDs up to
> >   apic_id_limit = x86_cpu_apic_id_from_index(max_cpus).
> > 
> > That leads to creation of MADT entries and Processor
> > objects in ACPI tables for not possible CPUs.
> > Fix it by allowing board specify a concrete list of
> > CPU IDs accourding its own rules (which for x86 depends
> > on topology). So that code that needs this list could
> > request it from board instead of trying to figure out
> > what IDs are correct on its own.
> > 
> > Signed-off-by: Igor Mammedov <address@hidden>
> > ---
> >  hw/i386/pc.c        | 16 ++++++++++++++++
> >  include/hw/boards.h | 18 ++++++++++++++++++
> >  2 files changed, 34 insertions(+)
> > 
> > diff --git a/hw/i386/pc.c b/hw/i386/pc.c
> > index 9227bde..548ec64 100644
> > --- a/hw/i386/pc.c
> > +++ b/hw/i386/pc.c
> > @@ -1931,6 +1931,21 @@ static unsigned pc_cpu_index_to_socket_id(unsigned 
> > cpu_index)
> >      return topo.pkg_id;
> >  }
> >  
> > +static GArray *pc_possible_cpu_arch_ids(void)
> > +{
> > +    int i;
> > +    GArray *list = g_array_new(FALSE, FALSE, sizeof(CPUArchId));
> > +
> > +    for (i = 0; i < max_cpus; i++) {
> > +        CPUArchId val;
> > +
> > +        val.arch_id = x86_cpu_apic_id_from_index(i);
> > +        val.cpu = qemu_get_cpu_by_arch_id(val.arch_id);
> > +        g_array_append_val(list, val);
> > +    }
> > +    return list;
> > +}  
> 
> You claim this version is linear, but I don't see any change from
> v1, except for a whitespace change.
> 
> qemu_get_cpu_by_arch_id() is O(smp_cpus). This calls
> qemu_get_cpu_by_arch_id() max_cpus times, so this is
> O(max_cpus*smp_cpus).
I'm sorry, I've fixed it for ACPI tables only but not here.
I've just posted and alternative patch for possible CPU
in CPU hotplug introspection context
 '[RFC] QMP: add query-hotpluggable-cpus'
which creates list in 2*O(N) and also makes access to list
more typesafe. Lets see how it goes and later I can respin
this series with that patch included.