Re: [RFC 4/4] memcg: NUMA background reclaim

[Balbir Singh <balbir-23VcF4HTsmIX0ybBhKVfKdBPR1lH4CV8@public.gmane.org>]

KAMEZAWA Hiroyuki wrote:
> One aspect of difference in reclaim logic between global lru and memcg is 
>  * global LRU triggers memory reclaim at memory shortage.
>  * memcg LRU triggers memory reclaim at excess of usage.
> 
> Then, global LRU _know_ which node we should start reclaim from.
>  * start from a node at memory shortage or
>  * start from a node where memory allocation is waiting
> 
> WRT memcg, it's difficult to find where we should start because
> there is no memory shortage and LRU is splitted.
> (But per-zone-LRU is definitely necessary for scalability.)
> 
> This patch tries to deteremine a node for starting recalim by checking
> ratio of inactive pages/active pages in a node. And trying to avoid starting
> from a node with relatively small usage.
> Better algorithm is welcome.
> 
> Singed-off-by: KAMEZAWA Hiruyuki <kamezawa.hiroyu-+CUm20s59erQFUHtdCDX3A@public.gmane.org>
> 
> Index: mm-2.6.26-rc2-mm1/mm/memcontrol.c
> ===================================================================
> --- mm-2.6.26-rc2-mm1.orig/mm/memcontrol.c
> +++ mm-2.6.26-rc2-mm1/mm/memcontrol.c
> @@ -578,7 +578,7 @@ retry:
>  	if (state == RES_OVER_LIMIT) {
>  		if (!(gfp_mask & __GFP_WAIT))
>  			goto out;
> -		if (try_to_free_mem_cgroup_pages(mem, gfp_mask))
> +		if (try_to_free_mem_cgroup_pages(mem, -1, gfp_mask))
>  			goto retry;
>  		/*
>  		 * try_to_free_mem_cgroup_pages() might not give us a
> @@ -801,7 +801,7 @@ int mem_cgroup_shrink_usage(struct mm_st
>  	rcu_read_unlock();
> 
>  	do {
> -		progress = try_to_free_mem_cgroup_pages(mem, gfp_mask);
> +		progress = try_to_free_mem_cgroup_pages(mem, -1, gfp_mask);
>  	} while (!progress && --retry);
> 
>  	if (!retry)
> @@ -814,7 +814,7 @@ static void mem_cgroup_drop_all_pages(st
>  {
>  	int progress;
>  	while (!res_counter_empty(&mem->res)) {
> -		progress = try_to_free_mem_cgroup_pages(mem,
> +		progress = try_to_free_mem_cgroup_pages(mem, -1,
>  					GFP_HIGHUSER_MOVABLE);
>  		if (!progress) /* we did as much as possible */
>  			break;
> @@ -912,6 +912,62 @@ out:
>  /*
>   * background reclaim daemon.
>   */
> +
> +#ifdef CONFIG_NUMA
> +/*
> + * Because memory controller's memory reclaim doesn't come from memory shortage,
> + * we cannot know which node should be reclaimed in an easy way.
> + * This routine select a node with inactive pages to be a node for starting
> + * scanning.
> + */
> +int __select_best_node(struct mem_cgroup *mem)
> +{
> +	int nid;
> +	int best_node = -1;
> +	unsigned long highest_inactive_ratio = 0;
> +	unsigned long active, inactive, inactive_ratio, total, threshold, flags;
> +	struct mem_cgroup_per_zone *mz;
> +	int zid;
> +
> +	/*
> +	 * When a node's memory usage is smaller than 
> +  	 * total_usage/num_of_node * 75%, we don't select the node
> +	 */
> +	total = mem->res.usage >> PAGE_SHIFT;
> +	threshold = (total / num_node_state(N_HIGH_MEMORY)) * 3 / 4;
> +
> +	/*
> +	 * See nodemask.h, N_HIGH_MEMORY means that a node has memory
> +	 * can be used for user's memory.(i.e. not means HIGHMEM).
> +	 */
> +	for_each_node_state(nid, N_HIGH_MEMORY) {
> +		active = 0;
> +		inactive = 0;
> +
> +		for (zid = 0; zid < MAX_NR_ZONES; zid++) {
> +			mz = mem_cgroup_zoneinfo(mem, nid, zid);
> +			spin_lock_irqsave(&mz->lru_lock, flags);
> +			active += MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE);
> +			inactive +=
> +				MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE);
> +			spin_unlock_irqrestore(&mz->lru_lock, flags);
> +		}
> +
> +		if (active + inactive < threshold)
> +			continue;
> +		inactive_ratio = (inactive * 100) / (active + 1);
> +		if (inactive_ratio > highest_inactive_ratio)
> +			best_node = nid;

Shouldn't we update highest_inactive_ration here?

> +	}
> +	return best_node;
> +}
> +#else
> +int __select_best_node(struct mem_cgroup *mem)
> +{
> +	return 0;
> +}
> +#endif
> +
>  static int mem_cgroup_reclaim_daemon(void *data)
>  {
>  	DEFINE_WAIT(wait);
> @@ -935,13 +991,9 @@ static int mem_cgroup_reclaim_daemon(voi
>  			continue;
>  		}
>  		finish_wait(&mem->daemon.waitq, &wait);
> -		/*
> -		 * memory resource controller doesn't see NUMA memory usage
> -		 * balancing, becasue we cannot know what balancing is good.
> -		 * TODO: some annotation or heuristics to detect which node
> -		 * we should start reclaim from.
> -		 */
> -		ret = try_to_free_mem_cgroup_pages(mem, GFP_HIGHUSER_MOVABLE);
> +
> +		ret = try_to_free_mem_cgroup_pages(mem,
> +				__select_best_node(mem), GFP_HIGHUSER_MOVABLE);
> 
>  		yield();
>  	}
> Index: mm-2.6.26-rc2-mm1/mm/vmscan.c
> ===================================================================
> --- mm-2.6.26-rc2-mm1.orig/mm/vmscan.c
> +++ mm-2.6.26-rc2-mm1/mm/vmscan.c
> @@ -1429,7 +1429,7 @@ unsigned long try_to_free_pages(struct z
>  #ifdef CONFIG_CGROUP_MEM_RES_CTLR
> 
>  unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
> -						gfp_t gfp_mask)
> +					   int nid, gfp_t gfp_mask)
>  {
>  	struct scan_control sc = {
>  		.may_writepage = !laptop_mode,
> @@ -1442,9 +1442,11 @@ unsigned long try_to_free_mem_cgroup_pag
>  	};
>  	struct zonelist *zonelist;
> 
> +	if (nid == -1)
> +		nid = numa_node_id();
>  	sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
>  			(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
> -	zonelist = NODE_DATA(numa_node_id())->node_zonelists;
> +	zonelist = NODE_DATA(nid)->node_zonelists;
>  	return do_try_to_free_pages(zonelist, &sc);
>  }
>  #endif
> Index: mm-2.6.26-rc2-mm1/include/linux/swap.h
> ===================================================================
> --- mm-2.6.26-rc2-mm1.orig/include/linux/swap.h
> +++ mm-2.6.26-rc2-mm1/include/linux/swap.h
> @@ -184,7 +184,7 @@ extern void swap_setup(void);
>  extern unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
>  					gfp_t gfp_mask);
>  extern unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem,
> -							gfp_t gfp_mask);
> +						int nid, gfp_t gfp_mask);
>  extern int __isolate_lru_page(struct page *page, int mode);
>  extern unsigned long shrink_all_memory(unsigned long nr_pages);
>  extern int vm_swappiness;
> 


-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL