Re: [Qemu-devel] [RFC PATCH v3] numa: enable sparse node numbering

[Nishanth Aravamudan <nacc@linux.vnet.ibm.com>]

On 25.06.2014 [13:52:56 -0300], Eduardo Habkost wrote:
> On Wed, Jun 25, 2014 at 09:13:59AM -0700, Nishanth Aravamudan wrote:
> > On 25.06.2014 [13:21:34 +0200], Igor Mammedov wrote:
> > > On Tue, 24 Jun 2014 10:40:38 -0700
> > > Nishanth Aravamudan <nacc@linux.vnet.ibm.com> wrote:
> > <snip>
> > > > diff --git a/include/sysemu/sysemu.h b/include/sysemu/sysemu.h
> > > > index 277230d..b90bf66 100644
> > > > --- a/include/sysemu/sysemu.h
> > > > +++ b/include/sysemu/sysemu.h
> > > > @@ -145,11 +145,13 @@ extern int mem_prealloc;
> > > >   */
> > > >  #define MAX_CPUMASK_BITS 255
> > > >  
> > > > -extern int nb_numa_nodes;
> > > > +extern int nb_numa_nodes; /* Number of NUMA nodes */
> > > > +extern int max_numa_node; /* Highest specified NUMA node ID */
> > > >  typedef struct node_info {
> > > >      uint64_t node_mem;
> > > >      DECLARE_BITMAP(node_cpu, MAX_CPUMASK_BITS);
> > > >      struct HostMemoryBackend *node_memdev;
> > > > +    bool present;
> > > How about dropping 'present' and replacing array with a list
> > > of only present nodes?
> > 
> > If that would be preferred, I can move to that. I assume a simple
> > linked-list is fine. Does qemu provide any infrastructure for defining
> > lists? I'll look through the source but any pointers would be helpful.
> > 
> > Generally speaking, sparse NUMA nodes aren't that common and when they
> > exist, the gaps aren't large. But it does seem to make sense if we have
> > sparse IDs at all, we might as well move to a list.
> > 
> > In any case, moving to the list means we'd have a nodeid as part of the
> > structure instead.
> > 
> > > That way it will be one more step closer to converting numa
> > > infrastructure to a set of QOM objects.
> > 
> > Sounds like a good idea to me. I'll respin the patch soon.
> 
> Having a list makes sense, the only difference is that keeping a sparse
> array sorted is much easier than making a sorted list (because the ACPI
> tables are nodeid-ordered). That's why I suggested keeping the array
> initially.

And for non-ACPI platforms, it does feel like keeping the list sorted is
ideal, as it simplifies various loops, etc.

> Adding a "present" field to the array is a trivial and easy-to-review
> change. Changing NodeInfo to use linked lists is a more complex change
> that I wouldn't want to include after soft freeze.
> 
> In other words:
>  * Having a list is better than a sparse array; but:
>  * Having a small sparse array with the "present" field is better
>    than broken sparse nodeid support (IMO).

Perhaps as a compromise I can work on the list conversion as a follow-on
patch?

Thanks,
Nish