On Fri, Mar 15, 2024 at 02:19:10PM +0800, Miaohe Lin wrote:
> On 2024/3/14 22:33, Matthew Wilcox wrote:
> > I actually want to do a bit more here ...
> >
> > +++ b/mm/debug.c
> > @@ -58,15 +58,10 @@ static void __dump_folio(struct folio *folio, struct page *page,
> > int mapcount = 0;
> > char *type = "";
> >
> > - /*
> > - * page->_mapcount space in struct page is used by slab pages to
> > - * encode own info, and we must avoid calling page_folio() again.
> > - */
> > - if (!folio_test_slab(folio)) {
> > + if (!page_has_type(page)) {
> > mapcount = atomic_read(&page->_mapcount) + 1;
> > - if (folio_test_large(folio))
> > - mapcount += folio_entire_mapcount(folio);
> > - }
> > + if (folio_test_large(folio))
> > + mapcount += folio_entire_mapcount(folio);
>
> Why folio_test_large is not within the "if (!page_has_type(page))" block? I think a slab page could also be large folio.
> Or am I miss something?
Slab pages don't use the first tail page so folio_entire_mapcount will
be 0 for them. And we don't want the folio_test_large() to be gated by
page_has_type() because hugetlb pages (after this patch) will be in the
page_has_type() category. So for hugetlb pages, we want to not read
page->mapcount (leaving mapcount at zero), but we do want to read the
entire_mapcount.
