Re: [PATCH] mm: fix pgalloc_stall on unpopulated zone

From: Minchan Kim
Date: Wed Jul 13 2016 - 21:10:12 EST

On Wed, Jul 13, 2016 at 10:25:04AM +0100, Mel Gorman wrote:
> On Wed, Jul 13, 2016 at 11:24:13AM +0900, Minchan Kim wrote:
> > If we use sc->reclaim_idx for accounting pgstall, it can increase
> > the count on unpopulated zone, for example, movable zone(but
> > my system doesn't have movable zone) if allocation request were
> > GFP_HIGHUSER_MOVABLE. It doesn't make no sense.
> >
> I wanted to track the highest zone allowed by each allocation regardless
> of what the zone population state was. Otherwise, consider the following
> on a NUMA system
> 1. An allocation request arrives for GFP_HIGHUSER_MOVABLE that stalls
> 2. System has two nodes, node 0 with ZONE_NORMAL, node 1 with ZONE_HIGHMEM
> 3. If the allocating process is on node 0, the stall is accounted on ZONE_NORMAL
> 4. If the allocatinn process is on node 1, the stall is accounted on ZONE_HIGHMEM
> Multiple runs of the same workload on the same machine will see stall
> statistics on different zones and renders the stat useless. This is
> difficult to analyse because stalls accounted for on ZONE_NORMAL may or
> may not be zone-constrained allocations.

Fair enough.
For the allocstall, it would be better to show the requested zone.

> The patch means that the vmstat accounting and tracepoint data is also
> out of sync. One thing I wanted to be able to do was
> 1. Observe that there are alloc stalls on DMA32 or some other low zone
> 2. Activate mm_vmscan_direct_reclaim_begin, filter on classzone_idx ==
> DMA32 and identify the source of the lowmem allocations
> If your patch is applied, I cannot depend on the stall stats any more
> and the tracepoint is required to determine if there really any
> zone-contrained allocations. It can be *inferred* from the skip stats
> but only if such skips occurred and that is not guaranteed.

Just a nit:

Hmm, can't we omit classzone_idx in mm_vm_scan_direct_begin_template?
Because every functions already have gfp_flags so that we can classzone_idx
via gfp_zone(gfp_flags) without passing it.

> --
> Mel Gorman
> SUSE Labs