On Wed, Mar 09, 2016 at 07:28:21AM +0300, Konstantin Khlebnikov wrote:
> On Tue, Mar 8, 2016 at 8:58 AM, Naoya Horiguchi
> <n-horigu...@ah.jp.nec.com> wrote:
> > On Tue, Mar 08, 2016 at 08:12:09AM +0300, Konstantin Khlebnikov wrote:
> >> On Tue, Mar 8, 2016 at 4:47 AM, Naoya Horiguchi
> >> <n-horigu...@ah.jp.nec.com> wrote:
> >> > I found that page-types is very slow and my testing shows many timeout
> >> > errors.
> >> > Here's an example with a simple program allocating 1000 thps.
> >> >
> >> > $ time ./page-types -p $(pgrep -f test_alloc)
> >> > ...
> >> > real 0m17.201s
> >> > user 0m16.889s
> >> > sys 0m0.312s
> >> >
> >> > $ time ./page-types.patched -p $(pgrep -f test_alloc)
> >> > ...
> >> > real 0m0.182s
> >> > user 0m0.046s
> >> > sys 0m0.135s
> >> >
> >> > Most of time is spent in memset(), which isn't necessary because we check
> >> > that the return of kpagecgroup_read() is equal to pages and uninitialized
> >> > memory is never used. So we can drop this memset().
> >>
> >> These zeros are used in show_page_range() - for merging pages into ranges.
> >
> > Hi Konstantin,
> >
> > Thank you for the response. The below code does solve the problem, so
> > that's fine.
> >
> > But I don't understand how the zeros are used. show_page_range() is called
> > via add_page() which is called for i=0 to i=pages-1, and the buffer cgi is
> > already filled for the range [i, pages-1] by kpagecgroup_read(), so even if
> > without zero initialization, kpagecgroup_read() properly fills zeros, right?
> > IOW, is there any problem if we don't do this zero initialization?
>
> kpagecgroup_read() reads only if kpagecgroup were opened,
> /proc/kpagecgroup might even not exist. Probably it's better to fill
> them with zeros here.
> Pre-memset was an optimization - it fills buffer only once instead on
> each kpagecgroup_read() call.
Ah, OK.
So here's ver.2.
Thanks,
Naoya
---
From: Naoya Horiguchi <n-horigu...@ah.jp.nec.com>
Subject: [PATCH v2] tools/vm/page-types.c: avoid memset() in walk_pfn() when
count == 1
I found that page-types is very slow and my testing shows many timeout errors.
Here's an example with a simple program allocating 1000 thps.
$ time ./page-types -p $(pgrep -f test_alloc)
...
real 0m17.201s
user 0m16.889s
sys 0m0.312s
Most of time is spent in memset(). Currently memset() clears over whole buffer
for every walk_pfn() call, which is inefficient when walk_pfn() is called from
walk_vma(), because in that case walk_pfn() is called for each pfn.
So this patch limits the zero initialization only for the first element.
$ time ./page-types.patched -p $(pgrep -f test_alloc)
...
real 0m0.182s
user 0m0.046s
sys 0m0.135s
Fixes: 954e95584579 ("tools/vm/page-types.c: add memory cgroup dumping and
filtering")
Signed-off-by: Naoya Horiguchi <n-horigu...@ah.jp.nec.com>
Suggested-by: Konstantin Khlebnikov <koc...@gmail.com>
---
tools/vm/page-types.c | 10 +++++++++-
1 file changed, 9 insertions(+), 1 deletion(-)
diff --git a/tools/vm/page-types.c b/tools/vm/page-types.c
index dab61c377f54..e92903fc7113 100644
--- a/tools/vm/page-types.c
+++ b/tools/vm/page-types.c
@@ -633,7 +633,15 @@ static void walk_pfn(unsigned long voffset,
unsigned long pages;
unsigned long i;
- memset(cgi, 0, sizeof cgi);
+ /*
+ * kpagecgroup_read() reads only if kpagecgroup were opened, but
+ * /proc/kpagecgroup might even not exist, so it's better to fill
+ * them with zeros here.
+ */
+ if (count == 1)
+ cgi[0] = 0;
+ else
+ memset(cgi, 0, sizeof cgi);
while (count) {
batch = min_t(unsigned long, count, KPAGEFLAGS_BATCH);
--
2.4.3
