Re: [PATCH 2/2] mm,memory_hotplug: {READ,WRITE}_ONCE unsynchronized zone data

[Lance Yang <ioworker0@gmail.com>]

On Wed, May 22, 2024 at 4:38 PM Brendan Jackman <jackmanb@google.com> wrote:
>
> Hi Lance, thanks for taking a look.
>
> On Wed, May 22, 2024 at 12:25:30PM +0800, Lance Yang wrote:
> > Hi Brendan,
> >
> > On Tue, May 21, 2024 at 8:57 PM Brendan Jackman <jackmanb@google.com> wrote:
> > > @@ -1077,7 +1081,7 @@ void adjust_present_page_count(struct page *page, struct memory_group *group,
> > >          */
> > >         if (early_section(__pfn_to_section(page_to_pfn(page))))
> > >                 zone->present_early_pages += nr_pages;
> > > -       zone->present_pages += nr_pages;
> > > +       WRITE_ONCE(zone->present_pages, zone->present_pages + nr_pages);
> >
> > I'm not sure that using the WRITE_ONCE() wrapper would prevent load tearing
> > on 'zone->present_pages', but it's probably just me overthinking it :)
>
> Hmm.. this isn't for load-tearing, it's for store-tearing. I have a
> feeling I might be missing your pont here though, can you elaborate?

Sorry, my explanation wasn't clear :(

I'm a bit confused about whether 'WRITE_ONCE(zone->present_pages,
zone->present_pages + nr_pages);'
is equivalent to the following:

1 a = zone->present_pages + nr_pages;
2 WRITE_ONCE(zone->present_pages, a);

If so, is there any possibility of load tearing on
'zone->present_pages' in line 1?

>
> I have just noticed that the original "big bad optimizing compiler"
> article[1] only says store-tearing has been observed in the wild when
> the value being stored can be split into immediates (i.e. is
> constant). But it doesn't really seem wise to rely on that. From what
> I can tell from tools/memory-model/Documentation you are really out in
> the wild with unmarked accesses.
>
> [1] https://lwn.net/Articles/793253

Thanks for clarifying!
Lance