Re: [PATCH] memcg: make mem_cgroup_read_stat() unsigned

[Andrew Morton <akpm-de/tnXTf+JLsfHDXvbKv3WD2FQJk+8+b@public.gmane.org>]

On Tue, 22 Sep 2015 17:42:13 -0700 Greg Thelen <gthelen-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org> wrote:

> Andrew Morton wrote:
> 
> > On Tue, 22 Sep 2015 15:16:32 -0700 Greg Thelen <gthelen-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org> wrote:
> >
> >> mem_cgroup_read_stat() returns a page count by summing per cpu page
> >> counters.  The summing is racy wrt. updates, so a transient negative sum
> >> is possible.  Callers don't want negative values:
> >> - mem_cgroup_wb_stats() doesn't want negative nr_dirty or nr_writeback.
> >> - oom reports and memory.stat shouldn't show confusing negative usage.
> >> - tree_usage() already avoids negatives.
> >>
> >> Avoid returning negative page counts from mem_cgroup_read_stat() and
> >> convert it to unsigned.
> >
> > Someone please remind me why this code doesn't use the existing
> > percpu_counter library which solved this problem years ago.
> >
> >>   for_each_possible_cpu(cpu)
> >
> > and which doesn't iterate across offlined CPUs.
> 
> I found [1] and [2] discussing memory layout differences between:
> a) existing memcg hand rolled per cpu arrays of counters
> vs
> b) array of generic percpu_counter
> The current approach was claimed to have lower memory overhead and
> better cache behavior.
> 
> I assume it's pretty straightforward to create generic
> percpu_counter_array routines which memcg could use.  Possibly something
> like this could be made general enough could be created to satisfy
> vmstat, but less clear.
> 
> [1] http://www.spinics.net/lists/cgroups/msg06216.html
> [2] https://lkml.org/lkml/2014/9/11/1057

That all sounds rather bogus to me.  __percpu_counter_add() doesn't
modify struct percpu_counter at all except for when the cpu-local
counter overflows the configured batch size.  And for the memcg
application I suspect we can set the batch size to INT_MAX...