On Tue, 2015-11-10 at 12:01 -0800, David Rientjes wrote:
> On Sat, 31 Oct 2015, Ben Hutchings wrote:
> 
> > If we need to increase the number of huge pages, drop caches first
> > to reduce fragmentation and then check that we actually allocated
> > as many as we wanted.  Retry once if that doesn't work.
> > 
> > Signed-off-by: Ben Hutchings <ben@decadent.org.uk>
> > ---
> > The test always fails for me in a 1 GB VM without this.
> > 
> > Ben.
> > 
> >  tools/testing/selftests/vm/run_vmtests | 15 ++++++++++++++-
> >  1 file changed, 14 insertions(+), 1 deletion(-)
> > 
> > diff --git a/tools/testing/selftests/vm/run_vmtests b/tools/testing/selftests/vm/run_vmtests
> > index 9179ce8..97ed1b2 100755
> > --- a/tools/testing/selftests/vm/run_vmtests
> > +++ b/tools/testing/selftests/vm/run_vmtests
> > @@ -20,13 +20,26 @@ done < /proc/meminfo
> >  if [ -n "$freepgs" ] && [ -n "$pgsize" ]; then
> >       nr_hugepgs=`cat /proc/sys/vm/nr_hugepages`
> >       needpgs=`expr $needmem / $pgsize`
> > -     if [ $freepgs -lt $needpgs ]; then
> > +     tries=2
> > +     while [ $tries -gt 0 ] && [ $freepgs -lt $needpgs ]; do
> >               lackpgs=$(( $needpgs - $freepgs ))
> > +             echo 3 > /proc/sys/vm/drop_caches
> >               echo $(( $lackpgs + $nr_hugepgs )) > /proc/sys/vm/nr_hugepages
> >               if [ $? -ne 0 ]; then
> >                       echo "Please run this test as root"
> >                       exit 1
> >               fi
> > +             while read name size unit; do
> > +                     if [ "$name" = "HugePages_Free:" ]; then
> > +                             freepgs=$size
> > +                     fi
> > +             done < /proc/meminfo
> > +             tries=$((tries - 1))
> > +     done
> > +     if [ $freepgs -lt $needpgs ]; then
> > +             printf "Not enough huge pages available (%d < %d)\n" \
> > +                    $freepgs $needpgs
> > +             exit 1
> >       fi
> >  else
> >       echo "no hugetlbfs support in kernel?"
> > 
> 
> I know this patch is in -mm and hasn't been merged by Linus yet, but I'm 
> wondering why the multiple /proc/sys/vm/drop_caches is helping?  Would it 
> simply suffice to put a sleep in there instead or is drop_caches actually 
> doing something useful a second time around?

Initially I just retried setting nr_hugepages up to 10 times, which
wasn't sufficient.  Then I added the drop_caches, and after that
setting nr_hugepages tended to worked first time so I reduced the retry
count.  It might not be necessary to retry at all.

Ben.

-- 
Ben Hutchings
All the simple programs have been written, and all the good names taken.