[lkml]   [2010]   [Nov]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH 6/6] memcg: make mem_cgroup_page_stat() return value unsigned
On Fri, Nov 12, 2010 at 12:41:15PM -0800, Greg Thelen wrote:
> >> mem_cgroup_page_stat() has changed so it never returns
> >> error so convert the return value to the traditional page
> >> count type (unsigned long).
> >
> > This changelog feels a bit beside the point.
> >
> > What's really interesting is that we now don't consider negative sums
> > to be invalid anymore, but just assume zero! There is a real
> > semantical change here.
> Prior to this patch series mem_cgroup_page_stat() returned a negative
> value (specifically -EINVAL) to indicate that the current task was in
> the root_cgroup and thus the per-cgroup usage and limit counter were
> invalid. Callers treated all negative values as an indication of
> root-cgroup message.
> Unfortunately there was another way that mem_cgroup_page_stat() could
> return a negative value even when current was not in the root cgroup.
> Negative sums were a possibility due to summing of unsynchronized
> per-cpu counters. These occasional negative sums would fool callers
> into thinking that the current task was in the root cgroup.
> Would adding this description to the commit message address your
> concerns?

I'd just describe that summing per-cpu counters is racy, that we can
end up with negative results, and the only sensible handling of that
is to assume zero.

> > That the return type can then be changed to unsigned long is a nice
> > follow-up cleanup that happens to be folded into this patch.
> Good point. I can separate the change into two sub-patches:
> 1. use zero for a min-value (as described above)
> 2. change return value to unsigned

Sounds good. You can just fold the previous patch (adjusting the
callsites) into 2, which should take care of the ordering problem.

 \ /
  Last update: 2010-11-19 12:41    [W:0.044 / U:2.544 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site