Re: [PATCH v10 3/6] mm: Introduce Reported pages

[Michael S. Tsirkin]

On Mon, Sep 23, 2019 at 08:28:00AM -0700, Alexander Duyck wrote:
> On Mon, Sep 23, 2019 at 8:00 AM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> >
> > On Mon, Sep 23, 2019 at 07:50:15AM -0700, Alexander Duyck wrote:
> > > > > +static inline void
> > > > > +page_reporting_reset_boundary(struct zone *zone, unsigned int order, int mt)
> > > > > +{
> > > > > +     int index;
> > > > > +
> > > > > +     if (order < PAGE_REPORTING_MIN_ORDER)
> > > > > +             return;
> > > > > +     if (!test_bit(ZONE_PAGE_REPORTING_ACTIVE, &zone->flags))
> > > > > +             return;
> > > > > +
> > > > > +     index = get_reporting_index(order, mt);
> > > > > +     reported_boundary[index] = &zone->free_area[order].free_list[mt];
> > > > > +}
> > > >
> > > > So this seems to be costly.
> > > > I'm guessing it's the access to flags:
> > > >
> > > >
> > > >         /* zone flags, see below */
> > > >         unsigned long           flags;
> > > >
> > > >         /* Primarily protects free_area */
> > > >         spinlock_t              lock;
> > > >
> > > >
> > > >
> > > > which is in the same cache line as the lock.
> > >
> > > I'm not sure what you mean by this being costly?
> >
> > I've just been wondering why does will it scale report a 1.5% regression
> > with this patch.
> 
> Are you talking about data you have collected from a test you have
> run, or the data I have run?

About the kernel test robot auto report that was sent recently.

> In the case of the data I have run I notice almost no difference as
> long as the pages are not actually being madvised. Once I turn on the
> madvise then I start seeing the regression, but almost all of that is
> due to page zeroing/faulting. There isn't expected to be a gain from
> this patchset until you start having guests dealing with memory
> overcommit on the host. Then at that point the patch set should start
> showing gains when the madvise bits are enabled in QEMU.
> 
> Also the test I have been running is a modified version of the
> page_fault1 test to specifically target transparent huge pages in
> order to make this test that much more difficult, the standard
> page_fault1 test wasn't showing much of anything since the overhead
> for breaking a 2M page into 512 4K pages and zeroing those
> individually in the guest  was essentially drowning out the effect of
> the patches themselves.
> 
> - Alex