Re: [PATCH v2 1/4] mm/hugetlb: fix the inconsistency of /proc/sys/vm/nr_huge_pages

[Mike Kravetz <mike.kravetz@oracle.com>]

On 08/10/23 09:34, David Hildenbrand wrote:
> On 10.08.23 02:17, Mike Kravetz wrote:
> > On 08/08/23 17:13, Xueshi Hu wrote:
> > > On 8/8/23 15:58, David Hildenbrand wrote:
> > > > On 08.08.23 04:28, Xueshi Hu wrote:
> > > > > On 8/7/23 23:15, David Hildenbrand wrote:
> > > > > > On 06.08.23 09:48, Xueshi Hu wrote:
> > 
> > The question is 'Should we change the /proc/sys/vm/nr_hugepages (and sysctl)
> > interfaces to be consistent with all the other read/show interfaces?
> > 
> > The argument for changing is that consistency is good.  Why have one interface
> > that is not like the others?
> > 
> > The reason for not changing is that this is the oldest interface.  The
> > information/interfaces originally available in /proc were created in /sys.
> > And, as mentioned in the documentation the /proc interfaces were kept
> > for backward compatibility.  Unfortunately, the meaning of nr_hugepages
> > was changed the /sys interfaces were created.  Sigh!!!
> 
> Indeed, they were designed to be different and to just leave the /proc
> interface alone.
> 

I am not sure if this was the 'design'.  The commit to add the sysfs interfaces
is a3437870160c from 2008.  There is no mention of changing the meaning of
nr_hugepages when read/displayed.

It matters not if this was by design.  It has been this way for 15 years and
has become the expected behavior.

> > 
> > In the thread mentioned above, I was in agreement with Hu about changing
> > /proc/sys/vm/nr_hugepages to be consistent with other read/show interfaces.
> > Now, I am not sure.
> 
> My take would be to just leave /proc/sys/vm/nr_hugepages alone. Maybe
> pr_warn_once() when the interface is used to guide people away from that
> legacy interface + clarify the docs.

Now, I tend to agree that not modifying /proc/sys/vm/nr_hugepages may be
the right thing to do.  I 'know' of a DB that makes extensive use of this
and the corresponding sysctl interface.  A pr_warn_once() may help, but I
still see the warning "Using mlock ulimits for SHM_HUGETLB is obsolete"
in system logs. :(

> Your call. :)

I REALLY would like it if all these interfaces were consistent and showed the
same information.  However, this inconsistency has been there for 15+ years.
And, I know of users making extensive use of /proc/sys/vm/nr_hugepages.

Hu, did you get a report of this inconsistency from a customer/end user?
Or, is this something you and other developers noticed?
-- 
Mike Kravetz