Re: [Qemu-devel] [PATCH v3 06/18] numa: mirror cpu to node mapping in MachineState::possible_cpus

[Igor Mammedov <imammedo@redhat.com>]

On Fri, 26 May 2017 12:46:25 -0300
Eduardo Habkost <ehabkost@redhat.com> wrote:

> On Wed, May 10, 2017 at 01:29:50PM +0200, Igor Mammedov wrote:
> [...]
> > diff --git a/hw/core/machine.c b/hw/core/machine.c
> > index 2482c63..420c8c4 100644
> > --- a/hw/core/machine.c
> > +++ b/hw/core/machine.c
> > @@ -389,6 +389,102 @@ HotpluggableCPUList *machine_query_hotpluggable_cpus(MachineState *machine)  
> [...]
> > +void machine_set_cpu_numa_node(MachineState *machine,
> > +                               const CpuInstanceProperties *props, Error **errp)
> > +{  
> [...]
> > +    /* force board to initialize possible_cpus if it hasn't been done yet */
> > +    mc->possible_cpu_arch_ids(machine);  
> [...]
> > diff --git a/numa.c b/numa.c
> > index 7182481..7db5dde 100644
> > --- a/numa.c
> > +++ b/numa.c
> > @@ -170,6 +170,7 @@ static void parse_numa_node(MachineState *ms, NumaNodeOptions *node,
> >          exit(1);
> >      }
> >      for (cpus = node->cpus; cpus; cpus = cpus->next) {
> > +        CpuInstanceProperties props;
> >          if (cpus->value >= max_cpus) {
> >              error_setg(errp,
> >                         "CPU index (%" PRIu16 ")"
> > @@ -178,6 +179,10 @@ static void parse_numa_node(MachineState *ms, NumaNodeOptions *node,
> >              return;
> >          }
> >          bitmap_set(numa_info[nodenr].node_cpu, cpus->value, 1);
> > +        props = mc->cpu_index_to_instance_props(ms, cpus->value);
> > +        props.node_id = nodenr;
> > +        props.has_node_id = true;
> > +        machine_set_cpu_numa_node(ms, &props, &error_fatal);  
> 
> This triggers a call to possible_cpu_arch_ids() before
> nb_numa_nodes is set to the actual number of NUMA nodes in the
> machine, breaking the "node_id = ... % nb_numa_nodes"
> initialization logic in pc, virt, and spapr.
> 
> The initialization ordering between possible_cpus and NUMA data
> structures looks very subtle and fragile.  I still don't see an
> obvious way to untangle that.
It's unfixable unless we require specific ordering on CLI,
i.e. first go all '-numa node,nodeid=[...]' options and only then
the rest of [-numa node,cpus|cpu].
We can do that for '-numa cpu' (probably should do enforce it for
this new option anyway for saner CLI)

but not for '-numa node,cpus' as it will break existing users
that do not declare nodes first.

> I suggest moving the default-NUMA-mapping code to a separate
> machine class method, instead of relying on
> possible_cpu_arch_ids() to initialize node_id.
So as you suggest we have to postpone default values initialization
till all the options are parsed:

1: strait-forward additional machine callback called from
   machine_run_board_init()

or:

2: save extra callback and recalculate not yet set props.node_id-s
   in possible_cpu_arch_ids() if nb_numa_nodes is changed since
   the last invocation of possible_cpu_arch_ids()

which one would you prefer?