Re: split_huge_page_to_list() races with page_mapcount() on migration entry in smaps code? [was: Re: [syzbot] kernel BUG in __page_mapcount]

[Kirill A. Shutemov]

On Mon, Jun 07, 2021 at 07:27:23PM +0200, Jann Horn wrote:
> 2. let smaps assume that the mapcount is 1 for all migration entries?

I believe that what we effectively do for migration entries to
non-compound pages:

	for (i = 0; i < nr; i++, page++) {
		int mapcount = page_mapcount(page);
		unsigned long pss = PAGE_SIZE << PSS_SHIFT;
		if (mapcount >= 2)
			pss /= mapcount;
		smaps_page_accumulate(mss, page, PAGE_SIZE, pss, dirty, locked,
				      mapcount < 2);
	}

For non-compound pages with page_count(page) != 1 (== 1 handled
separately) we would have nr == 1 and will look into mapcount, which for
pages under migration is 0. The code above will handle mapcount == 0 as
mapcount == 1. I think it would not be a stretch to do the same for
compound pages here.

I guess we should take an additional argument to smaps_account() which would
indicate that we deal with migration entry and handle it as mapcount == 1.

Hm. Do we need the same for device-private entries?

-- 
 Kirill A. Shutemov