Re: [PATCH 06/27] mm, vmscan: Begin reclaiming pages on a per-node basis

[Johannes Weiner]

On Tue, Feb 23, 2016 at 03:04:29PM +0000, Mel Gorman wrote:
> @@ -2428,10 +2448,11 @@ static bool shrink_zone(struct zone *zone, struct scan_control *sc,
>  			reclaimed = sc->nr_reclaimed;
>  			scanned = sc->nr_scanned;
>  
> +			sc->reclaim_idx = reclaim_idx;
>  			shrink_zone_memcg(zone, memcg, sc, &lru_pages);
>  			zone_lru_pages += lru_pages;

The setting of sc->reclaim_idx is unexpected here. Why not set it in
the caller and eliminate the reclaim_idx parameter?

> @@ -2558,16 +2579,12 @@ static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
>  		sc->gfp_mask |= __GFP_HIGHMEM;
>  
>  	for_each_zone_zonelist_nodemask(zone, z, zonelist,
> -					requested_highidx, sc->nodemask) {

It's unfortunate that we start with the lowest zone here. For Normal
allocations, the most common allocations, this will always have two
full shrink_node() rounds that skip over everything >DMA in the first,
then over everything >DMA32 in the second, even though all pages on
the node are valid reclaim candidates for that allocation.