Re: [PATCH -next 1/7] mm_types: add _last_cpupid into folio

[Kefeng Wang]

On 2023/10/11 13:55, Huang, Ying wrote:
> Kefeng Wang <wangkefeng.wang@xxxxxxxxxx> writes:
>
> > On 2023/10/10 20:33, Matthew Wilcox wrote:
> > > On Tue, Oct 10, 2023 at 02:45:38PM +0800, Kefeng Wang wrote:
> > > > At present, only arc/sparc/m68k define WANT_PAGE_VIRTUAL, both of
> > > > them don't support numa balancing, and the page struct is aligned
> > > > to _struct_page_alignment, it is safe to move _last_cpupid before
> > > > 'virtual' in page, meanwhile, add it into folio, which make us to
> > > > use folio->_last_cpupid directly.
> > > What do you mean by "safe"?  I think you mean "Does not increase the
> > > size of struct page", but if that is what you mean, why not just say so?
> > > If there's something else you mean, please explain.
> >
> > Don't increase size of struct page and don't impact the real order of
> > struct page as the above three archs without numa balancing support.
> >
> > > In any event, I'd like to see some reasoning that _last_cpupid is
> > > actually
> > > information which is logically maintained on a per-allocation basis,
> > > not a per-page basis (I think this is true, but I honestly don't know)
> >
> > The _last_cpupid is updated in should_numa_migrate_memory() from numa
> > fault(do_numa_page, and do_huge_pmd_numa_page), it is per-page(normal
> > page and PMD-mapped page). Maybe I misunderstand your mean, please
> > correct me.
>
> Because PTE mapped THP will not be migrated according to comments and
> folio_test_large() test in do_numa_page().  Only _last_cpuid of the head
> page will be used (that is, on per-allocation basis).  Although in
> change_pte_range() in mprotect.c, _last_cpuid of tail pages may be
> changed, they are not used actually.  All in all, _last_cpuid is on
> per-allocation basis for now.

Thanks for clarification, yes, it's what I mean, too
> In the future, it's hard to say.  PTE-mapped THPs or large folios give
> us an opportunity to check whether the different parts of a folio are
> accessed by multiple sockets, so that we should split the folio.  But
> this is just some possibility in the future.

It depends on memory access behavior of application,if multiple sockets
access a large folio/PTE-mappped THP frequently, split maybe better,
or it is enough to just migrate the entire folio.


> --
> Best Regards,
> Huang, Ying