On 02/13/23 19:01, Michal Hocko wrote:
On Fri 03-02-23 12:16:04, Mike Kravetz wrote: [...]
Unless someone thinks we should move forward, I will not push the code for this approach now. It will also be interesting to see if this is impacted at all by the outcome of discussions to perhaps redesign mapcount.
Yes, I do agree. We might want to extend page_mapcount documentation a bit though. The comment is explicit about the order-0 pages but a note about hugetlb and pmd sharing wouldn't hurt. WDYT?
Looks like that comment about 'Mapcount of 0-order page' has been removed in the latest version of page_mapcount(). It would not surprise me if the calls to page_mapcount after which we check for shared PMDs will soon be replaced with calls to folio_mapcount().
Perhaps Matthew has an opinion as to where map counts for hugetlb shared PMDs might be mentioned.