Re: [PATCH v3 1/2] numa: Require distance map when empty node exists

[Andrew Jones <drjones@redhat.com>]

On Wed, Oct 13, 2021 at 01:53:46PM +0200, Igor Mammedov wrote:
> On Wed, 13 Oct 2021 13:35:44 +0200
> Andrew Jones <drjones@redhat.com> wrote:
> 
> > On Wed, Oct 13, 2021 at 01:30:11PM +0200, Igor Mammedov wrote:
> > > On Wed, 13 Oct 2021 12:58:04 +0800
> > > Gavin Shan <gshan@redhat.com> wrote:
> > >   
> > > > The following option is used to specify the distance map. It's
> > > > possible the option isn't provided by user. In this case, the
> > > > distance map isn't populated and exposed to platform. On the
> > > > other hand, the empty NUMA node, where no memory resides, is
> > > > allowed on platforms like ARM64 virt. For these empty NUMA
> > > > nodes, their corresponding device-tree nodes aren't populated,
> > > > but their NUMA IDs should be included in the "/distance-map"
> > > > device-tree node, so that kernel can probe them properly if
> > > > device-tree is used.
> > > > 
> > > >   -numa,dist,src=<numa_id>,dst=<numa_id>,val=<distance>
> > > > 
> > > > This adds extra check after the machine is initialized, to
> > > > ask for the distance map from user when empty nodes exist in
> > > > device-tree.
> > > > 
> > > > Signed-off-by: Gavin Shan <gshan@redhat.com>
> > > > ---
> > > >  hw/core/machine.c     |  4 ++++
> > > >  hw/core/numa.c        | 24 ++++++++++++++++++++++++
> > > >  include/sysemu/numa.h |  1 +
> > > >  3 files changed, 29 insertions(+)
> > > > 
> > > > diff --git a/hw/core/machine.c b/hw/core/machine.c
> > > > index b8d95eec32..c0765ad973 100644
> > > > --- a/hw/core/machine.c
> > > > +++ b/hw/core/machine.c
> > > > @@ -1355,6 +1355,10 @@ void machine_run_board_init(MachineState *machine)
> > > >      accel_init_interfaces(ACCEL_GET_CLASS(machine->accelerator));
> > > >      machine_class->init(machine);
> > > >      phase_advance(PHASE_MACHINE_INITIALIZED);
> > > > +
> > > > +    if (machine->numa_state) {
> > > > +        numa_complete_validation(machine);
> > > > +    }
> > > >  }
> > > >  
> > > >  static NotifierList machine_init_done_notifiers =
> > > > diff --git a/hw/core/numa.c b/hw/core/numa.c
> > > > index 510d096a88..7404b7dd38 100644
> > > > --- a/hw/core/numa.c
> > > > +++ b/hw/core/numa.c
> > > > @@ -727,6 +727,30 @@ void numa_complete_configuration(MachineState *ms)
> > > >      }
> > > >  }
> > > >  
> > > > +/*
> > > > + * When device-tree is used by the machine, the empty node IDs should
> > > > + * be included in the distance map. So we need provide pairs of distances
> > > > + * in this case.
> > > > + */
> > > > +void numa_complete_validation(MachineState *ms)
> > > > +{
> > > > +    NodeInfo *numa_info = ms->numa_state->nodes;
> > > > +    int nb_numa_nodes = ms->numa_state->num_nodes;
> > > > +    int i;
> > > > +
> > > > +    if (!ms->fdt || ms->numa_state->have_numa_distance) {  
> > > 
> > > also skip check/limitation when VM is launched with ACPI enabled?  
> > 
> > Even with ACPI enabled we generate a DT and would like that DT to be as
> > complete as possible. Of course we should generate a SLIT table with
> 
> Guest will work just fine without distance map as SRAT describes
> all numa nodes.
> You are forcing VM to have SLIT just for the sake of 'completeness'
> that's practically unused.
> 
> I'm still unsure about pushing all of this in generic numa code,
> as this will be used only by ARM for now. It's better to keep it
> ARM specific, and when RISCV machine will start using this, it
> could be moved to generic code.

I don't see a problem in providing this DT node / ACPI table to guests
with empty NUMA nodes. I don't even see a problem with providing the
distance information unconditionally. Can you explain why we should
prefer not to provide optional HW descriptions?

Thanks,
drew 

> 
> > the distance information the user provides on the command line in order
> > to satisfy the check, and we will, since we already have that code in
> > place.
> 
> 
> > 
> > Thanks,
> > drew
> > 
> > >   
> > > > +        return;
> > > > +    }
> > > > +
> > > > +    for (i = 0; i < nb_numa_nodes; i++) {
> > > > +        if (numa_info[i].present && !numa_info[i].node_mem) {
> > > > +            error_report("Empty node %d found, please provide "
> > > > +                         "distance map.", i);
> > > > +            exit(EXIT_FAILURE);
> > > > +        }
> > > > +    }
> > > > +}
> > > > +
> > > >  void parse_numa_opts(MachineState *ms)
> > > >  {
> > > >      qemu_opts_foreach(qemu_find_opts("numa"), parse_numa, ms, &error_fatal);
> > > > diff --git a/include/sysemu/numa.h b/include/sysemu/numa.h
> > > > index 4173ef2afa..80f25ab830 100644
> > > > --- a/include/sysemu/numa.h
> > > > +++ b/include/sysemu/numa.h
> > > > @@ -104,6 +104,7 @@ void parse_numa_hmat_lb(NumaState *numa_state, NumaHmatLBOptions *node,
> > > >  void parse_numa_hmat_cache(MachineState *ms, NumaHmatCacheOptions *node,
> > > >                             Error **errp);
> > > >  void numa_complete_configuration(MachineState *ms);
> > > > +void numa_complete_validation(MachineState *ms);
> > > >  void query_numa_node_mem(NumaNodeMem node_mem[], MachineState *ms);
> > > >  extern QemuOptsList qemu_numa_opts;
> > > >  void numa_cpu_pre_plug(const struct CPUArchId *slot, DeviceState *dev,  
> > >   
> > 
>