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

[Mel Gorman <mgorman@techsingularity.net>]

On Tue, Dec 20, 2016 at 03:35:02PM +0100, Michal Hocko wrote:
> On Tue 20-12-16 14:28:45, Mel Gorman wrote:
> > On Tue, Dec 20, 2016 at 02:26:43PM +0100, Michal Hocko wrote:
> > > On Tue 20-12-16 13:10:40, Mel Gorman wrote:
> > > > On Tue, Dec 20, 2016 at 10:18:14AM +0100, 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.
> > > > > 
> > > > 
> > > > This is a similar concern to what I had. If the preferred zone, which is
> > > > the first valid usable zone, is not a "hit" for the statistics then I
> > > > don't know what "hit" is meant to mean.
> > > 
> > > But the first valid usable zone is defined based on the requested numa
> > > node. Unless the requested node is memoryless then we should have a hit,
> > > no?
> > > 
> > 
> > Should be. If the local node is memoryless then there would be a difference
> > between hit and whether it's local or not but that to me is a little
> > useless. A local vs remote page allocated has a specific meaning and
> > consequence. It's hard to see how hit can be meaningfully interpreted if
> > there are memoryless nodes. I don't have a strong objection to the patch
> > so I didn't nak it, I'm just not convinced it matters.
> 
> So what do you think about
> http://lkml.kernel.org/r/20161220091814.GC3769@dhcp22.suse.cz
> 

This doesn't appear to resolve for me and I've 30 minutes left before
being offline for 4 days so didn't go digging. 

> I think that we should get rid of __GFP_OTHER_NODE thingy. It is just
> one off thing and the gfp space it rather precious.
> 

However, broadly speaking, I'd be ok with getting rid of
__GFP_OTHER_NODE altogether and making it truely only about local vs
remote hits because those are the ones that matter in terms of
performance. If a user has memoryless nodes or policies that allow local
CPUs but forbid local memory and they need to debug an issue, they're
going to need tracepoints anyway. Hit/miss/other is not sufficient for
most interesting problems involving local or remote memory usage.

-- 
Mel Gorman
SUSE Labs