Re: mmotm 2009-09-09-22-56 uploaded

[Yinghai Lu <yinghai@kernel.org>]

David Rientjes wrote:
> On Fri, 18 Sep 2009, David Rientjes wrote:
> 
>> This seems to be related to 2547089 "x86/PCI: initialize PCI bus node 
>> numbers early" since cpumask_of_pcibus() on x86 doesn't check for -1 
>> pcibus_to_node() like most other architectures.  It'll simply index into 
>> cpumask_of_node for whatever the pci_sysdata's node is, and in this case 
>> that's -1.
>>
> 
> Suresh, could you give this patch a try?  It turns out that the 
> mp_bus_to_node map simply leaves all busses that don't have memory 
> affinity to -1, so cpu_online_mask is actually the appropriate cpumask to 
> return and x86 doesn't catch this.
> 
> 
> x86: default pcibus cpumask to all cpus if it lacks affinity
> 
> The early initialization of the pci bus to node mapping leaves all busses
> with a node id of -1 if it lacks memory affinity.  Thus, cpumask_of_pcibus
> must return all online cpus for such busses.
> 
> Signed-off-by: David Rientjes <rientjes@google.com>
> ---
>  arch/x86/include/asm/pci.h |    6 +++++-
>  1 files changed, 5 insertions(+), 1 deletions(-)
> 
> diff --git a/arch/x86/include/asm/pci.h b/arch/x86/include/asm/pci.h
> --- a/arch/x86/include/asm/pci.h
> +++ b/arch/x86/include/asm/pci.h
> @@ -143,7 +143,11 @@ static inline int __pcibus_to_node(const struct pci_bus *bus)
>  static inline const struct cpumask *
>  cpumask_of_pcibus(const struct pci_bus *bus)
>  {
> -	return cpumask_of_node(__pcibus_to_node(bus));
> +	int node;
> +
> +	node = __pcibus_to_node(bus);
> +	return (node == -1) ? cpu_online_mask :
> +			      cpumask_of_node(node);
>  }
>  #endif
>  

looks right, but we should use nearby_node or local cpu node instead of cpu_online_mask...

YH