Re: [Qemu-arm] [PATCH 2/3] numa: move default mapping init to machine

[Igor Mammedov]

On Thu, 18 May 2017 15:50:58 -0300
Eduardo Habkost <address@hidden> wrote:

> On Thu, May 18, 2017 at 10:09:30AM +0200, Igor Mammedov wrote:
> > there is no need use cpu_index_to_instance_props() for setting
> > default cpu -> node mapping. Generic machine code can do it
> > without cpu_index by just enabling already preset defaults
> > in possible_cpus.
> > 
> > PS:
> > as bonus it makes one less user of cpu_index_to_instance_props()
> > 
> > Signed-off-by: Igor Mammedov <address@hidden>
> > ---
> >  hw/core/machine.c | 32 +++++++++++++++++++++-----------
> >  numa.c            | 26 --------------------------
> >  2 files changed, 21 insertions(+), 37 deletions(-)
> > 
> > diff --git a/hw/core/machine.c b/hw/core/machine.c
> > index fd6a436..2e91aa9 100644
> > --- a/hw/core/machine.c
> > +++ b/hw/core/machine.c
> > @@ -700,26 +700,36 @@ static char *cpu_slot_to_string(const CPUArchId *cpu)
> >      return g_string_free(s, false);
> >  }
> >  
> > -static void machine_numa_validate(MachineState *machine)
> > +static void machine_numa_finish_init(MachineState *machine)
> >  {
> > -    int i;
> > +    int i, default_mapping;  
> 
> I suggest bool instead of int.
will do it in v2

> 
> >      GString *s = g_string_new(NULL);
> >      MachineClass *mc = MACHINE_GET_CLASS(machine);
> >      const CPUArchIdList *possible_cpus = 
> > mc->possible_cpu_arch_ids(machine);
> >  
> >      assert(nb_numa_nodes);
> > +    for (i = possible_cpus->len;
> > +         i && !possible_cpus->cpus[i - 1].props.has_node_id;
> > +         i--)
> > +        ;;  
> 
> I believe the original code was more readable, and it had only 1
> more line than this version. i.e.:
> 
>     for (i = 0; i < possible_cpus->len; i++) {
>         if (possible_cpus->cpus[i].props.has_node_id) {
>             break;
>         }
>     }
>     default_mapping = (i == possible_cpus->len);
ok, I'll do it this way

> 
> > +    default_mapping = !i; /* i == 0 : no explicit mapping provided by user 
> > */
> > +
> >      for (i = 0; i < possible_cpus->len; i++) {
> >          const CPUArchId *cpu_slot = &possible_cpus->cpus[i];
> >  
> > -        /* at this point numa mappings are initilized by CLI options
> > -         * or with default mappings so it's sufficient to list
> > -         * all not yet mapped CPUs here */
> > -        /* TODO: make it hard error in future */  
> 
> Did we change our mind about making it a hard error in the
> future?
nope, error message at the end of function says that partial mapping
is obsoleted and will be removed. I've thought that's was sufficient
reminder for us of what needs to be removed.
I'll move TODO to:

+            } else {
+                /* record slots with not set mapping */
++               /* TODO: make it hard error in future */ 
+                char *cpu_str = cpu_slot_to_string(cpu_slot);
+                g_string_append_printf(s, "%sCPU %d [%s]",

> 
> >          if (!cpu_slot->props.has_node_id) {
> > -            char *cpu_str = cpu_slot_to_string(cpu_slot);
> > -            g_string_append_printf(s, "%sCPU %d [%s]", s->len ? ", " : "", 
> > i,
> > -                                   cpu_str);
> > -            g_free(cpu_str);
> > +            if (default_mapping) {
> > +                /* fetch default mapping from board and enable it */
> > +                CpuInstanceProperties props = cpu_slot->props;
> > +                props.has_node_id = true;
> > +                machine_set_cpu_numa_node(machine, &props, &error_fatal);  
> 
> Is a machine_set_cpu_numa_node() call really necessary here, if
> we are already looking at cpu_slot->props directly?
yes, it's necessary as cpu_slot comes from:
   const CPUArchIdList *possible_cpus = mc->possible_cpu_arch_ids(machine);
callback, which initializes machine::possible_cpus if necessary.

So cpu_slot is readonly and I'd prefer to keep mc->possible_cpu_arch_ids()
returning 'const' as it's used by external callers and they shouldn't
mess with possible_cpus content by accident.

usage of machine_set_cpu_numa_node() adds +2 more lines and it's
proper/validating setter for node_id and will catch
mistakes if we make them in the code.
So I wouldn't use shortcuts here to save 2 lines.

 
> > +            } else {
> > +                /* record slots with not set mapping */
> > +                char *cpu_str = cpu_slot_to_string(cpu_slot);
> > +                g_string_append_printf(s, "%sCPU %d [%s]",
> > +                                       s->len ? ", " : "", i, cpu_str);
> > +                g_free(cpu_str);
> > +            }
> >          }  
> 
> What about doing this instead:
> 
>         if (default_mapping) {
>             /*
>              * Default mapping was already set by board at
>              * cpu_slot->props.node_id, just enable it
>              */
>             cpu_slot->props.has_node_id = true;
>         } else if (!cpu_slot->props.has_node_id) {
>             char *cpu_str = cpu_slot_to_string(cpu_slot);
>             g_string_append_printf(s, "%sCPU %d [%s]", s->len ? ", " : "", i,
>                                    cpu_str);
>             g_free(cpu_str);
>         }