On Fri, Nov 23, 2018 at 09:19:00AM -0800, Matthew Wilcox wrote:
> On Fri, Nov 23, 2018 at 01:56:44PM +0300, Kirill A. Shutemov wrote:
> > On Thu, Nov 22, 2018 at 01:32:24PM -0800, Matthew Wilcox wrote:
> > > Transparent Huge Pages are currently stored in i_pages as pointers to
> > > consecutive subpages. This patch changes that to storing consecutive
> > > pointers to the head page in preparation for storing huge pages more
> > > efficiently in i_pages.
> >
> > I probably miss something, I don't see how it wouldn't break
> > split_huge_page().
> >
> > I don't see what would replace head pages in i_pages with
> > formerly-tail-pages?
>
> You're quite right. Where's your test-suite? ;-)
Yeah-yeah...
> I think this should do the job:
>
> +++ b/mm/huge_memory.c
> @@ -2464,6 +2464,9 @@ static void __split_huge_page(struct page *page, struct list_head *list,
> if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
> shmem_uncharge(head->mapping->host, 1);
> put_page(head + i);
> + } else if (!PageAnon(page)) {
> + __xa_store(&head->mapping->i_pages, head[i].index,
> + head + i, 0);
> }
> }
Looks good to me. But I still need to look into the rest of the patch.
> Having looked at this area, I think there was actually a bug in the patch
> you wrote that I'm cribbing from. You inserted the tail pages before
> calling __split_huge_page_tail(), so a racing lookup would have found
> a tail page before it got transformed into a non-tail page.
I don't think so.
The page still has refcount==0 and any lookup of the page suppose to fail
due to !page_cache_get_speculative() or block on tree lock.
--
Kirill A. Shutemov
