Re: [PATCH RFC 1/1] mm, page_alloc: fix incorrect zone_statistics data

[hejianet <hejianet@gmail.com>]

On 20/12/2016 5:18 PM, Michal Hocko wrote:
> On Mon 12-12-16 13:59:07, Jia He wrote:
>> In commit b9f00e147f27 ("mm, page_alloc: reduce branches in
>> zone_statistics"), it reconstructed codes to reduce the branch miss rate.
>> Compared with the original logic, it assumed if !(flag & __GFP_OTHER_NODE)
>>   z->node would not be equal to preferred_zone->node. That seems to be
>> incorrect.
> I am sorry but I have hard time following the changelog. It is clear
> that you are trying to fix a missed NUMA_{HIT,OTHER} accounting
> but it is not really clear when such thing happens. You are adding
> preferred_zone->node check. preferred_zone is the first zone in the
> requested zonelist. So for the most allocations it is a node from the
> local node. But if something request an explicit numa node (without
> __GFP_OTHER_NODE which would be the majority I suspect) then we could
> indeed end up accounting that as a NUMA_MISS, NUMA_FOREIGN so the
> referenced patch indeed caused an unintended change of accounting AFAIU.
>
> If this is correct then it should be a part of the changelog. I also
> cannot say I would like the fix. First of all I am not sure
> __GFP_OTHER_NODE is a good idea at all. How is an explicit usage of the
> flag any different from an explicit __alloc_pages_node(non_local_nid)?
> In both cases we ask for an allocation on a remote node and successful
> allocation is a NUMA_HIT and NUMA_OTHER.
>
> That being said, why cannot we simply do the following? As a bonus, we
> can get rid of a barely used __GFP_OTHER_NODE. Also the number of
> branches will stay same.
Yes, I agree maybe we can get rid of __GFP_OTHER_NODE if no objections
Seems currently it is only used for hugepage and statistics
> ---
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 429855be6ec9..f035d5c8b864 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -2583,25 +2583,17 @@ int __isolate_free_page(struct page *page, unsigned int order)
>    * Update NUMA hit/miss statistics
>    *
>    * Must be called with interrupts disabled.
> - *
> - * When __GFP_OTHER_NODE is set assume the node of the preferred
> - * zone is the local node. This is useful for daemons who allocate
> - * memory on behalf of other processes.
>    */
>   static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
>   								gfp_t flags)
>   {
>   #ifdef CONFIG_NUMA
> -	int local_nid = numa_node_id();
> -	enum zone_stat_item local_stat = NUMA_LOCAL;
> -
> -	if (unlikely(flags & __GFP_OTHER_NODE)) {
> -		local_stat = NUMA_OTHER;
> -		local_nid = preferred_zone->node;
> -	}
> +	if (z->node == preferred_zone->node) {
> +		enum zone_stat_item local_stat = NUMA_LOCAL;
>   
> -	if (z->node == local_nid) {
>   		__inc_zone_state(z, NUMA_HIT);
> +		if (z->node != numa_node_id())
> +			local_stat = NUMA_OTHER;
>   		__inc_zone_state(z, local_stat);
>   	} else {
>   		__inc_zone_state(z, NUMA_MISS);
I thought the logic here is different
Here is the zone_statistics() before introducing __GFP_OTHER_NODE:

if (z->zone_pgdat == preferred_zone->zone_pgdat) {
         __inc_zone_state(z, NUMA_HIT);
     } else {
         __inc_zone_state(z, NUMA_MISS);
         __inc_zone_state(preferred_zone, NUMA_FOREIGN);
     }
     if (z->node == numa_node_id())
         __inc_zone_state(z, NUMA_LOCAL);
     else
         __inc_zone_state(z, NUMA_OTHER);

B.R.
Jia