Re: [PATCH v2 4/8] mm: migrate: use a folio in migrate_misplaced_page()

[Kefeng Wang]

On 2023/8/29 10:10, Matthew Wilcox wrote:
> On Tue, Aug 29, 2023 at 08:49:31AM +0800, Huang, Ying wrote:
> > Zi Yan <ziy@xxxxxxxxxx> writes:
> >
> > > On 21 Aug 2023, at 7:56, Kefeng Wang wrote:
> > >
> > > > Use a folio in migrate_misplaced_page() to save compound_head() calls.
> > > >
> > > > Signed-off-by: Kefeng Wang <wangkefeng.wang@xxxxxxxxxx>
> > > > ---
> > > >   mm/migrate.c | 23 ++++++++++++-----------
> > > >   1 file changed, 12 insertions(+), 11 deletions(-)
> > >
> > > LGTM. And a comment below. Reveiwed-by: Zi Yan <ziy@xxxxxxxxxx>
> > >
> > > > diff --git a/mm/migrate.c b/mm/migrate.c
> > > > index 281eafdf8e63..fc728f9a383f 100644
> > > > --- a/mm/migrate.c
> > > > +++ b/mm/migrate.c
> > > > @@ -2521,17 +2521,18 @@ int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
> > > >   			   int node)
> > > >   {
> > > >   	pg_data_t *pgdat = NODE_DATA(node);
> > > > +	struct folio *folio = page_folio(page);
> > > >   	int isolated;
> > > >   	int nr_remaining;
> > > >   	unsigned int nr_succeeded;
> > > >   	LIST_HEAD(migratepages);
> > > > -	int nr_pages = thp_nr_pages(page);
> > > > +	int nr_pages = folio_nr_pages(folio);
> > > >
> > > >   	/*
> > > >   	 * Don't migrate file pages that are mapped in multiple processes
> > > >   	 * with execute permissions as they are probably shared libraries.
> > > >   	 */
> > > > -	if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
> > > > +	if (page_mapcount(page) != 1 && folio_is_file_lru(folio) &&
> > >
> > > page_mapcount() is not converted, since folio_mapcount() is not equivalent
> > > to page_mapcount(). It can be converted and this function can be converted
> > > to migrate_misplaced_folio() once we have something like folio_num_sharers().
> >
> > It seems that we can use folio_estimated_sharers() here.
>
> So, funny thing, page_mapcount() was always wrong here.  We have two
> callers, do_huge_pmd_numa_page() and do_numa_page().  do_numa_page()
> has a check for PageCompound() (and /* TODO: handle PTE-mapped THP */).
> do_huge_pmd_numa_page() returns pfn_to_page(), after it got the pfn
> fromm pmd_pfn(pmd).

for PMD-mapped page，do_huge_pmd_numa_page() return head page

   unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
   page = vm_normal_page_pmd(vma, haddr, pmd);

do_numa_page() only handle base page with PageCompound() check， so we
could use folio_estimated_sharers()， I am not clear why page_mapcount
is wrong here, could you explain more, thank.

> That makes folio_estimated_sharers() an improvement, possibly even
> a bugfix.  Also, we should look at removing the PageCompound() check
> in do_numa_page().

Yes, for pte-mapped thp and the large folio numa migrate support, we
need remove the check, we could do it in another patchset.

>