Re: [Patch v2] mm/rmap.c: split huge pmd when it really is

[Kirill A. Shutemov]

On Fri, Jan 03, 2020 at 10:01:28PM +0800, Wei Yang wrote:
> On Fri, Jan 03, 2020 at 04:26:50PM +0300, Kirill A. Shutemov wrote:
> >On Fri, Jan 03, 2020 at 09:05:54PM +0800, Wei Yang wrote:
> >> On Fri, Jan 03, 2020 at 03:18:46PM +0800, Wei Yang wrote:
> >> >On Tue, Dec 24, 2019 at 06:28:56AM +0800, Wei Yang wrote:
> >> >>When page is not NULL, function is called by try_to_unmap_one() with
> >> >>TTU_SPLIT_HUGE_PMD set. There are two cases to call try_to_unmap_one()
> >> >>with TTU_SPLIT_HUGE_PMD set:
> >> >>
> >> >>  * unmap_page()
> >> >>  * shrink_page_list()
> >> >>
> >> >>In both case, the page passed to try_to_unmap_one() is PageHead() of the
> >> >>THP. If this page's mapping address in process is not HPAGE_PMD_SIZE
> >> >>aligned, this means the THP is not mapped as PMD THP in this process.
> >> >>This could happen when we do mremap() a PMD size range to an un-aligned
> >> >>address.
> >> >>
> >> >>Currently, this case is handled by following check in __split_huge_pmd()
> >> >>luckily.
> >> >>
> >> >>  page != pmd_page(*pmd)
> >> >>
> >> >>This patch checks the address to skip some work.
> >> >
> >> >I am sorry to forget address Kirill's comment in 1st version.
> >> >
> >> >The first one is the performance difference after this change for a PTE
> >> >mappged THP.
> >> >
> >> >Here is the result:(in cycle)
> >> >
> >> >        Before     Patched
> >> >
> >> >        963        195
> >> >        988        40
> >> >        895        78
> >> >
> >> >Average 948        104
> >> >
> >> >So the change reduced 90% time for function split_huge_pmd_address().
> >
> >Right.
> >
> >But do we have a scenario, where the performance of
> >split_huge_pmd_address() matters? I mean, it it called as part of rmap
> >walk, attempt to split huge PMD where we don't have huge PMD should be
> >within noise.
> 
> Sorry for my poor English.
> 
> I don't catch the meaning of the last sentence. "within noise" here means
> non-huge PMD is an expected scenario and we could tolerate this? 

Basically, I doubt that this change would bring any measurable perfromance
benefits on a real workload.

-- 
 Kirill A. Shutemov