On Fri, Dec 11, 2020 at 3:23 PM Jason Gunthorpe <jgg@xxxxxxxx> wrote:
>
> On Fri, Dec 11, 2020 at 03:21:39PM -0500, Pavel Tatashin wrote:
> > @@ -1593,7 +1592,7 @@ static long check_and_migrate_cma_pages(struct mm_struct *mm,
> > }
> >
> > if (!isolate_lru_page(head)) {
> > - list_add_tail(&head->lru, &cma_page_list);
> > + list_add_tail(&head->lru, &movable_page_list);
> > mod_node_page_state(page_pgdat(head),
> > NR_ISOLATED_ANON +
> > page_is_file_lru(head),
> > @@ -1605,7 +1604,7 @@ static long check_and_migrate_cma_pages(struct mm_struct *mm,
> > i += step;
> > }
> >
> > - if (!list_empty(&cma_page_list)) {
> > + if (!list_empty(&movable_page_list)) {
>
> You didn't answer my earlier question, is it OK that ZONE_MOVABLE
> pages leak out here if ioslate_lru_page() fails but the
> moval_page_list is empty?
>
> I think the answer is no, right?
In my opinion it is OK. We are doing our best to not pin movable
pages, but if isolate_lru_page() fails because pages are currently
locked by someone else, we will end up long-term pinning them.
See comment in this patch:
+ * 1. Pinned pages: (long-term) pinning of movable pages is avoided
+ * when pages are pinned and faulted, but it is still possible that
+ * address space already has pages in ZONE_MOVABLE at the time when
+ * pages are pinned (i.e. user has touches that memory before
+ * pinning). In such case we try to migrate them to a different zone,
+ * but if migration fails the pages can still end-up pinned in
+ * ZONE_MOVABLE. In such case, memory offlining might retry a long
+ * time and will only succeed once user application unpins pages.
>
> Jason
