lkml.org 
[lkml]   [2016]   [Jun]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC PATCH 2/3] mm, thp: convert from optimistic to conservative
Hello,

On Sat, Jun 11, 2016 at 10:16:00PM +0300, Ebru Akagunduz wrote:
> Currently, khugepaged collapses pages saying only
> a referenced page enough to create a THP.
>
> This patch changes the design from optimistic to conservative.
> It gives a default threshold which is half of HPAGE_PMD_NR
> for referenced pages, also introduces a new sysfs knob.

Strictly speaking, It's not what I suggested.

I didn't mean that let's change threshold for deciding whether we should
collapse or not(although just *a* reference page seems be too
optimistic) and export the knob to the user. In fact, I cannot judge
whether it's worth or not because I never have an experience with THP
workload in practice although I believe it does make sense.

What I suggested is that a swapin operation would be much heavier than
a THP cost to collapse populated anon page so it should be more
conservative than THP collasping decision, at least. Given that thought,
decision point for collasping a THP is *a* reference page now so *half*
reference of populated pages for reading swapped-out page is more
conservative.

>
> Signed-off-by: Ebru Akagunduz <ebru.akagunduz@gmail.com>
> ---
> include/trace/events/huge_memory.h | 10 ++++----
> mm/khugepaged.c | 50 +++++++++++++++++++++++++++++---------
> 2 files changed, 44 insertions(+), 16 deletions(-)
>
> diff --git a/include/trace/events/huge_memory.h b/include/trace/events/huge_memory.h
> index 830d47d..5f14025 100644
> --- a/include/trace/events/huge_memory.h
> +++ b/include/trace/events/huge_memory.h
> @@ -13,7 +13,7 @@
> EM( SCAN_EXCEED_NONE_PTE, "exceed_none_pte") \
> EM( SCAN_PTE_NON_PRESENT, "pte_non_present") \
> EM( SCAN_PAGE_RO, "no_writable_page") \
> - EM( SCAN_NO_REFERENCED_PAGE, "no_referenced_page") \
> + EM( SCAN_LACK_REFERENCED_PAGE, "lack_referenced_page") \
> EM( SCAN_PAGE_NULL, "page_null") \
> EM( SCAN_SCAN_ABORT, "scan_aborted") \
> EM( SCAN_PAGE_COUNT, "not_suitable_page_count") \
> @@ -47,7 +47,7 @@ SCAN_STATUS
> TRACE_EVENT(mm_khugepaged_scan_pmd,
>
> TP_PROTO(struct mm_struct *mm, struct page *page, bool writable,
> - bool referenced, int none_or_zero, int status, int unmapped),
> + int referenced, int none_or_zero, int status, int unmapped),
>
> TP_ARGS(mm, page, writable, referenced, none_or_zero, status, unmapped),
>
> @@ -55,7 +55,7 @@ TRACE_EVENT(mm_khugepaged_scan_pmd,
> __field(struct mm_struct *, mm)
> __field(unsigned long, pfn)
> __field(bool, writable)
> - __field(bool, referenced)
> + __field(int, referenced)
> __field(int, none_or_zero)
> __field(int, status)
> __field(int, unmapped)
> @@ -108,14 +108,14 @@ TRACE_EVENT(mm_collapse_huge_page,
> TRACE_EVENT(mm_collapse_huge_page_isolate,
>
> TP_PROTO(struct page *page, int none_or_zero,
> - bool referenced, bool writable, int status),
> + int referenced, bool writable, int status),
>
> TP_ARGS(page, none_or_zero, referenced, writable, status),
>
> TP_STRUCT__entry(
> __field(unsigned long, pfn)
> __field(int, none_or_zero)
> - __field(bool, referenced)
> + __field(int, referenced)
> __field(bool, writable)
> __field(int, status)
> ),
> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> index e3d8da7..43fc41e 100644
> --- a/mm/khugepaged.c
> +++ b/mm/khugepaged.c
> @@ -27,7 +27,7 @@ enum scan_result {
> SCAN_EXCEED_NONE_PTE,
> SCAN_PTE_NON_PRESENT,
> SCAN_PAGE_RO,
> - SCAN_NO_REFERENCED_PAGE,
> + SCAN_LACK_REFERENCED_PAGE,
> SCAN_PAGE_NULL,
> SCAN_SCAN_ABORT,
> SCAN_PAGE_COUNT,
> @@ -68,6 +68,7 @@ static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
> */
> static unsigned int khugepaged_max_ptes_none __read_mostly;
> static unsigned int khugepaged_max_ptes_swap __read_mostly;
> +static unsigned int khugepaged_min_ptes_young __read_mostly;

We should set it to 1 to preserve old behavior.

\
 
 \ /
  Last update: 2016-06-15 09:21    [W:0.090 / U:0.516 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site