On Thu, 3 Jul 2014 17:37:29 +0900
Minchan Kim <minchan@xxxxxxxxxx> wrote:
> Hello,
>
> On Thu, Jul 03, 2014 at 10:29:01AM +0200, Martin Schwidefsky wrote:
> > On Thu, 3 Jul 2014 16:29:54 +0900
> > Minchan Kim <minchan@xxxxxxxxxx> wrote:
> >
> > > Hello,
> > >
> > > On Thu, Jul 03, 2014 at 10:03:19AM +0900, Minchan Kim wrote:
> > > > Hello,
> > > >
> > > > On Tue, Jul 01, 2014 at 05:50:58PM +0300, Kirill A. Shutemov wrote:
> > > > > On Tue, Jul 01, 2014 at 09:36:15AM +0900, Minchan Kim wrote:
> > > > > > + do {
> > > > > > + /*
> > > > > > + * XXX: We can optimize with supporting Hugepage free
> > > > > > + * if the range covers.
> > > > > > + */
> > > > > > + next = pmd_addr_end(addr, end);
> > > > > > + if (pmd_trans_huge(*pmd))
> > > > > > + split_huge_page_pmd(vma, addr, pmd);
> > > > >
> > > > > Could you implement proper THP support before upstreaming the feature?
> > > > > It shouldn't be a big deal.
> > > >
> > > > Okay, Hope to review.
> > > >
> > > > Thanks for the feedback!
> > > >
> > >
> > > I tried to implement it but had a issue.
> > >
> > > I need pmd_mkold, pmd_mkclean for MADV_FREE operation and pmd_dirty for
> > > page_referenced. When I investigate all of arches supported THP,
> > > it's not a big deal but s390 is not sure to me who has no idea of
> > > soft tracking of s390 by storage key instead of page table information.
> > > Cced s390 maintainer. Hope to help.
> >
> > Storage key for dirty and referenced tracking is a thing of the past.
> > The current code for s390 uses software tracking for dirty and referenced.
> > There is one catch though, for ptes the software implementation covers
> > dirty and referenced bit but for pmds only referenced bit is available.
> > The reason is that there is no free bit left in the pmd entry for the
> > software dirty bit.
>
> Thanks for the quick reply.
>
> >
> > > So, if there isn't any help from s390, I should introduce
> > > HAVE_ARCH_THP_MADVFREE to disable MADV_FREE support of THP in s390 but
> > > not want to introduce such new config.
> >
> > Why is the dirty bit for pmds needed for the MADV_FREE implementation?
>
> MADV_FREE semantic want it.
>
> When madvise syscall is called, VM clears dirty bit of ptes of
> the range. If memory pressure happens, VM checks dirty bit of
> page table and if it found still "clean", it means it's a
> "lazyfree pages" so VM could discard the page instead of swapping out.
> Once there was store operation for the page before VM peek a page
> to reclaim, dirty bit is set so VM can swap out the page instead of
> discarding to keep up-to-date contents.
>
> If it's hard on s390, maybe we could use just reference bit
> instead of dirty bit to check recent access but it might change
> semantic a bit with other OSes. :(
Just discussed this with Gerald and we found a trick how we can add
a dirty bit to the pmd entries. That will be a non-trivial patch but
we can do it. Until that time you could just define pmd_dirty to
always return true and the code should "work" in the sense that it
does not break anything.
--
blue skies,
Martin.
"Reality continues to ruin my life." - Calvin.
--
To unsubscribe from this list: send the line "unsubscribe linux-api" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
