Re: [RESEND RFC PATCH 2/2] gfp: use the best near online node if the target node is offline

[Gu Zheng <guz.fnst@cn.fujitsu.com>]

Hi Andrew,

On 04/25/2015 04:01 AM, Andrew Morton wrote:

> On Fri, 24 Apr 2015 17:58:33 +0800 Gu Zheng <guz.fnst@cn.fujitsu.com> wrote:
> 
>> Since the change to the cpu <--> mapping (map the cpu to the physical
>> node for all possible at the boot), the node of cpu may be not present,
>> so we use the best near online node if the node is not present in the low
>> level allocation APIs.
>>
>> ...
>>
>> --- a/include/linux/gfp.h
>> +++ b/include/linux/gfp.h
>> @@ -298,9 +298,31 @@ __alloc_pages(gfp_t gfp_mask, unsigned int order,
>>  	return __alloc_pages_nodemask(gfp_mask, order, zonelist, NULL);
>>  }
>>  
>> +static int find_near_online_node(int node)
>> +{
>> +	int n, val;
>> +	int min_val = INT_MAX;
>> +	int best_node = -1;
>> +
>> +	for_each_online_node(n) {
>> +		val = node_distance(node, n);
>> +
>> +		if (val < min_val) {
>> +			min_val = val;
>> +			best_node = n;
>> +		}
>> +	}
>> +
>> +	return best_node;
>> +}
> 
> This should be `inline' if it's in a header file.
> 
> But it is far too large to be inlined anyway - please move it to a .c file.

Agree.

> 
> And please document it.  A critical thing to describe is how we
> determine whether a node is "near".  There are presumably multiple ways
> in which we could decide that a node is "near" (number of hops, minimum
> latency, ...).  Which one did you choose, and why?

It just reuse the dropped code in PATCH 1/2, based on the node_distance table,
which is a arch special defined one, and the data mostly comes from the
firmware info, e.g. SLIT table.

> 
>>  static inline struct page *alloc_pages_node(int nid, gfp_t gfp_mask,
>>  						unsigned int order)
>>  {
>> +	/* Offline node, use the best near online node */
>> +	if (!node_online(nid))
>> +		nid = find_near_online_node(nid);
>> +
>>  	/* Unknown node is current node */
>>  	if (nid < 0)
>>  		nid = numa_node_id();
>> @@ -311,7 +333,11 @@ static inline struct page *alloc_pages_node(int nid, gfp_t gfp_mask,
>>  static inline struct page *alloc_pages_exact_node(int nid, gfp_t gfp_mask,
>>  						unsigned int order)
>>  {
>> -	VM_BUG_ON(nid < 0 || nid >= MAX_NUMNODES || !node_online(nid));
>> +	/* Offline node, use the best near online node */
>> +	if (!node_online(nid))
>> +		nid = find_near_online_node(nid);
>> +
>> +	VM_BUG_ON(nid < 0 || nid >= MAX_NUMNODES);
>>  
>>  	return __alloc_pages(gfp_mask, order, node_zonelist(nid, gfp_mask));
>>  }
> 
> Ouch.  These functions are called very frequently, and adding overhead
> to them is a big deal.  And the patch even adds overhead to non-x86
> architectures which don't benefit from it!
> 
> Is there no way this problem can be fixed somewhere else?  Preferably
> by fixing things up at hotplug time.

As Kame suggested, maintaining a per-cpu cache about the alternative-node
only for x86 arch seems a good choice.

Regards,
Gu

> .
>