Messages in this thread | | | Date | Sat, 19 Mar 2016 04:01:01 +0300 | From | "Kirill A. Shutemov" <> | Subject | Re: [PATCHv4 04/25] rmap: support file thp |
| |
On Fri, Mar 18, 2016 at 03:10:06PM +0530, Aneesh Kumar K.V wrote: > "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> writes: > > > [ text/plain ] > > Naive approach: on mapping/unmapping the page as compound we update > > ->_mapcount on each 4k page. That's not efficient, but it's not obvious > > how we can optimize this. We can look into optimization later. > > > > PG_double_map optimization doesn't work for file pages since lifecycle > > of file pages is different comparing to anon pages: file page can be > > mapped again at any time. > > > > Can you explain this more ?. We added PG_double_map so that we can keep > page_remove_rmap simpler. So if it isn't a compound page we still can do > > if (!atomic_add_negative(-1, &page->_mapcount)) > > I am trying to understand why we can't use that with file pages ?
The first thing: for non-compound pages we still have simple atomic_inc_and_test() / atomic_add_negative(-1), nothing changed here.
About compound pages:
For anon-THP PG_double_map allowed to not touch _mapcount in all subpages until a PMD which maps the page is split. This way we significantly lower overhead on refcounting as long as we have the page mapped with PMD-only, since we only need to increment compound_mapcount().
The optimization is possible due to relatively simple lifecycle of anonymous THP page:
- anon-THPs always mapped with PMD first;
- new mapping of THP can only be created via fork();
- the page only can get mapped with PTEs via split_huge_pmd();
For file-THP the situation is different. Once we allocated a huge page and put it on radix tree, the page can be mapped with PTEs or PMDs at any time. It makes the same optimization inapplicable there.
I think there *can* be some room for optimization, but I don't want to invest more time here, until it's identified as bottleneck. It can lead to more complex code on rmap side.
-- Kirill A. Shutemov
| |