On Fri, Jan 13, 2023 at 09:16:23AM -0600, Sidhartha Kumar wrote:
> On 1/10/23 6:00 PM, Mike Kravetz wrote:
> > On 01/10/23 21:40, Matthew Wilcox wrote:
> > > On Tue, Jan 10, 2023 at 03:28:21PM -0600, Sidhartha Kumar wrote:
> > > > @@ -3505,6 +3505,7 @@ static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
> > > > {
> > > > int nr_nodes, node;
> > > > struct page *page;
> > > > + struct folio *folio;
> > > > lockdep_assert_held(&hugetlb_lock);
> > > > @@ -3518,8 +3519,8 @@ static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
> > > > list_for_each_entry(page, &h->hugepage_freelists[node], lru) {
> > > > if (PageHWPoison(page))
> > > > continue;
> > > > -
> > > > - return demote_free_huge_page(h, page);
> > > > + folio = page_folio(page);
> > > > + return demote_free_hugetlb_folio(h, folio);
> > > > }
> > > > }
> > >
> > > Can't this be
> > > list_for_each_entry(folio, &h->hugepage_freelists[node], lru)
> > >
> > > which avoids the call to page_folio() here.
> > >
> > > I think the call to PageHWPoison is actually wrong here. That would
> > > only check the hwpoison bit on the first page, whereas we want to know
> > > about the hwpoison bit on any page (don't we?) So this should be
> > > folio_test_has_hwpoisoned()?
> > >
> > > Or is that a THP-thing that is different for hugetlb pages?
> >
> > I believe it is different for hugetlb pages. See hugetlb_set_page_hwpoison()
> > where it sets PageHWPoison on head page as well as allocating a raw_hwp_page
>
> I agree, this line in hugetlb_set_page_hwpoison (which is now
> folio_set_hugetlb_hwpoison) sets the HWPoison flag on the head page
>
> int ret = folio_test_set_hwpoison(folio) ? -EHWPOISON : 0;
>
> so the correct code in demote_pool_huge_page() would be:
>
> list_for_each_entry(folio, &h->hugepage_freelists[node], lru) {
> if (folio_test_hwpoison(folio))
> continue;
> return demote_free_hugetlb_folio(h, folio);
> }
>
>
> > to track the actual page with poison. Note that we can not directly flag
> > hugetlb 'subpages' because we may not have the struct pages due to vmemmap
> > optimization. Adding Naoya just to be sure.
Yes, the above code and the reason totally make sense to me.
Thanks for the patch and the update.
- Naoya Horiguchif
