Re: [RFC PATCH 09/14] mm: Handle large folios in free_unref_folios()

[Matthew Wilcox <willy@infradead.org>]

On Thu, Aug 31, 2023 at 04:21:53PM +0100, Ryan Roberts wrote:
> On 25/08/2023 14:59, Matthew Wilcox (Oracle) wrote:
> > @@ -2478,7 +2478,11 @@ void free_unref_folios(struct folio_batch *folios)
> >  	for (i = 0, j = 0; i < folios->nr; i++) {
> >  		struct folio *folio = folios->folios[i];
> >  		unsigned long pfn = folio_pfn(folio);
> > -		if (!free_unref_page_prepare(&folio->page, pfn, 0))
> > +		unsigned int order = folio_order(folio);
> 
> Do you need to do anything special for hugetlb folios? I see that
> destroy_large_folio() has:
> 
> 	if (folio_test_hugetlb(folio)) {
> 		free_huge_folio(folio);
> 		return;
> 	}

Right; hugetlb folios get freed specially and never come this way.
I could put in an assertion, I suppose?

> > @@ -2486,11 +2490,13 @@ void free_unref_folios(struct folio_batch *folios)
> >  		 * comment in free_unref_page.
> >  		 */
> >  		migratetype = get_pcppage_migratetype(&folio->page);
> > -		if (unlikely(is_migrate_isolate(migratetype))) {
> > +		if (order > PAGE_ALLOC_COSTLY_ORDER ||
> 
> Should this be `if (!pcp_allowed_order(order) ||` ? That helper includes the THP
> pageblock_order too.

Oh, yes, that's obviously far better than what I had here.  I got the
BUG in order_to_pindex() and didn't think to look around for the
correct predicate.  Thanks!