On Mon 07-12-20 22:52:30, Muchun Song wrote:
> On Mon, Dec 7, 2020 at 9:00 PM Michal Hocko <mhocko@xxxxxxxx> wrote:
> >
> > On Sun 06-12-20 18:14:39, Muchun Song wrote:
> > > Hi,
> > >
> > > This patch series is aimed to convert all THP vmstat counters to pages
> > > and some KiB vmstat counters to bytes.
> > >
> > > The unit of some vmstat counters are pages, some are bytes, some are
> > > HPAGE_PMD_NR, and some are KiB. When we want to expose these vmstat
> > > counters to the userspace, we have to know the unit of the vmstat counters
> > > is which one. It makes the code complex. Because there are too many choices,
> > > the probability of making a mistake will be greater.
> > >
> > > For example, the below is some bug fix:
> > > - 7de2e9f195b9 ("mm: memcontrol: correct the NR_ANON_THPS counter of hierarchical memcg")
> > > - not committed(it is the first commit in this series) ("mm: memcontrol: fix NR_ANON_THPS account")
> > >
> > > This patch series can make the code simple (161 insertions(+), 187 deletions(-)).
> > > And make the unit of the vmstat counters are either pages or bytes. Fewer choices
> > > means lower probability of making mistakes :).
> > >
> > > This was inspired by Johannes and Roman. Thanks to them.
> >
> > It would be really great if you could summarize the current and after
> > the patch state so that exceptions are clear and easier to review. The
>
> Agree. Will do in the next version. Thanks.
>
>
> > existing situation is rather convoluted but we have at least units part
> > of the name so it is not too hard to notice that. Reducing exeptions
> > sounds nice but I am not really sure it is such an improvement it is
> > worth a lot of code churn. Especially when it comes to KB vs B. Counting
>
> There are two vmstat counters (NR_KERNEL_STACK_KB and
> NR_KERNEL_SCS_KB) whose units are KB. If we do this, all
> vmstat counter units are either pages or bytes in the end. When
> we expose those counters to userspace, it can be easy. You can
> reference to:
>
> [RESEND PATCH v2 11/12] mm: memcontrol: make the slab calculation consistent
>
> From this point of view, I think that it is worth doing this. Right?
Well, unless I am missing something, we have two counters in bytes, two
in kB, both clearly distinguishable by the B/KB suffix. Changing KB to B
will certainly reduce the different classes of units, no question about
that, but I am not really sure this is worth all the code churn. Maybe
others will think otherwise.
As I've said the THP accounting change makes more sense to me because it
allows future changes which are already undergoing so there is more
merit in those.
--
Michal Hocko
SUSE Labs
