Re: [PATCH] x86/numa: Adjust datatypes for node and pxm

[Boris Ostrovsky <boris.ostrovsky@oracle.com>]

On 02/23/2015 09:47 AM, Jan Beulich wrote:
>>>> On 23.02.15 at 15:42, <boris.ostrovsky@oracle.com> wrote:
>> On 02/23/2015 04:57 AM, Jan Beulich wrote:
>>>>>> On 21.02.15 at 19:14, <boris.ostrovsky@oracle.com> wrote:
>>>> --- a/xen/arch/x86/srat.c
>>>> +++ b/xen/arch/x86/srat.c
>>>> @@ -21,44 +21,55 @@
>>>>    #include <asm/e820.h>
>>>>    #include <asm/page.h>
>>>>    
>>>> +#define MAX_PXM   255
>>> Perhaps better (MAX_NUMNODES - 1) than a literal number? Or
>>> even do away with it altogether, use MAX_NUMNODES - 1 in the
>>> array definition and ARRAY_SIZE() elsewhere?
>> I am not sure we can do this: PXMs are not guaranteed to be zero-based.
>>
>> And, in fact, the way we map PXMs to nodes is not quite right because of
>> this --- it just so happens that PXMs are usually under 255 (and
>> zero-based) and we can use them as index to 255-sized array. IIRC the
>> spec defines them as 32-bit values.
> Ah, yes, you're right. We may need to allocate the array dynamically
> as soon as we run into a system with large PXMs.
>

But then we may end up with a huge and very sparse array. Maybe a hash 
with each element as a linked list? We will hit the first element in 
pretty much all cases so it should be reasonably fast.

-boris