Re: [PATCH 1/5] x86: Change size of node ids from u8 to u16 fixup

[Mike Travis <travis@sgi.com>]

David Rientjes wrote:
> On Sat, 19 Jan 2008, Mike Travis wrote:
> 
>>> Yeah, NID_INVAL is negative so no unsigned type will work here, 
>>> unfortunately.  And that reduces the intended savings of your change since 
>>> the smaller type can only be used with a smaller CONFIG_NODES_SHIFT.
>>>
>> Excuse my ignorance but why wouldn't this work:
>>
>> static numanode_t pxm_to_node_map[MAX_PXM_DOMAINS]
>>                                 = { [0 ... MAX_PXM_DOMAINS - 1] = NUMA_NO_NODE };
>> ...
>>>> int acpi_map_pxm_to_node(int pxm)
>>>> {
>>>         int node = pxm_to_node_map[pxm];
>>>
>>>         if (node < 0)
>> 	   numanode_t node = pxm_to_node_map[pxm];
>>
> 
> Because NUMA_NO_NODE is 0xff on x86.  That's a valid node id for 
> configurations with CONFIG_NODES_SHIFT equal to or greater than 8.

Perhaps numanode_t should be set to u16 if MAX_NUMNODES > 255 to
allow for an invalid value of 255? 

#if MAX_NUMNODES > 255
typedef u16 numanode_t;
#else
typedef u8 numanode_t;
#endif

> 
>> 	   if (node != NUMA_NO_NODE) {
> 
> Wrong, this should be
> 
> 	node == NUMA_NO_NODE

Oops, yes you're right.

>>>>                 if (nodes_weight(nodes_found_map) >= MAX_NUMNODES)
>>>>                         return NID_INVAL;
>>>>                 node = first_unset_node(nodes_found_map);
>>>>                 __acpi_map_pxm_to_node(pxm, node);
>>>>                 node_set(node, nodes_found_map);
>>>>         }
> 
> The net result of this is that if a proximity domain is looked up through 
> acpi_map_pxm_to_node() and already has a mapping to node 255 (legal with 
> CONFIG_NODES_SHIFT == 8), this function will return NID_INVAL since the 
> weight of nodes_found_map is equal to MAX_NUMNODES.

> 
> You simply can't use valid node id's to signify invalid or unused node 
> ids.
> 
>> or change:
>> 	#define NID_INVAL       (-1)
>> to
>> 	#define NID_INVAL       ((numanode_t)(-1))
>> ...
>> 	   if (node != NID_INVAL) {
> 
> You mean
> 
> 	node == NID_INVAL
> 
>>>>                 if (nodes_weight(nodes_found_map) >= MAX_NUMNODES)
>>>>                         return NID_INVAL;
>>>>                 node = first_unset_node(nodes_found_map);
>>>>                 __acpi_map_pxm_to_node(pxm, node);
>>>>                 node_set(node, nodes_found_map);
>>>>         }
> 
> That's the equivalent of your NUMA_NO_NODE code above.  The fact remains 
> that (numanode_t)-1 is still a valid node id for MAX_NUMNODES >= 256.
> 
> So, as I said in my initial reply, the only way to get the savings you're 
> looking for is to use u8 for CONFIG_NODES_SHIFT <= 7 and then convert all 
> NID_INVAL users to use NUMA_NO_NODE.

Yes, I agree.  I'll do the changes you're suggesting.

> Additionally, Linux has always discouraged typedefs when they do not 
> define an architecture-specific size.  The savings from your patch for 
> CONFIG_NODES_SHIFT == 7 would be 256 bytes for this mapping.
> 
> It's simply not worth it.

So are you saying that I should just use u16 for all node ids whether
CONFIG_NODES_SHIFT > 7 or not?  Othersise, I would think that defining a
typedef is a fairly clean solution.

A quick grep shows that there are 35 arrays defined by MAX_NUMNODES in
x86_64, 38 in X86_32 (not verified.)  So it's not exactly a trivial
amount of memory.

> 
>> And btw, shouldn't the pxm value be sized to numanode_t size as well?
>> Will it ever be larger than the largest node id?
>>
> 
> Section 6.2.9 of ACPI 2.0 states that PXM's return an integer, so that 
> would be non-conforming to the standard.
> 
> Additionally, PXM's are not nodes, so casting them to anything called 
> numanode_t shows the semantic flaw in your patch.

Thanks for the info.  I wasn't sure exactly what the PXM value represents.
> 
> 		David

Thanks again,
Mike