Date:   Thu, 20 Sep 2018 17:21:20 +0530
From:   Pintu Kumar <pintu.ping@...il.com>
To:     rppt@...ux.vnet.ibm.com
Cc:     open list <linux-kernel@...r.kernel.org>,
        Russell King - ARM Linux <linux@...linux.org.uk>,
        linux-arm-kernel@...ts.infradead.org, linux-mm@...ck.org
Subject: Re: KSM not working in 4.9 Kernel

Hi,

Thank you so much for all your reply so far.
I have few more doubts to understand the output from ksm sysfs.
Device: Hikey620 - ARM64 - Linux 4.9.20
With HUGE page enabled:
CONFIG_TRANSPARENT_HUGEPAGE=y
CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS=y
# CONFIG_TRANSPARENT_HUGEPAGE_MADVISE is not set

Currently, I get this output, when I run below program with ksm:

~ # grep -H '' /sys/kernel/mm/ksm/*
/sys/kernel/mm/ksm/full_scans:29
/sys/kernel/mm/ksm/page_comparisons:39584
/sys/kernel/mm/ksm/pages_hashed:11672
/sys/kernel/mm/ksm/pages_scanned:21766
/sys/kernel/mm/ksm/pages_shared:3
/sys/kernel/mm/ksm/pages_sharing:10097
/sys/kernel/mm/ksm/pages_to_scan:200
/sys/kernel/mm/ksm/pages_unshared:53
/sys/kernel/mm/ksm/pages_volatile:1
/sys/kernel/mm/ksm/run:0
/sys/kernel/mm/ksm/sleep_millisecs:1000
---------------------------

int main(int argc, char *argv[])
{
        int i, n, size, ret;
        char *buffer;
        void *addr;

        n = 100;
        size = 100 * getpagesize();
        for (i = 0; i < n; i++) {
                buffer = (char *)malloc(size);
                memset(buffer, 0xff, size);
                madvise(buffer, size, MADV_MERGEABLE);
                if (ret < 0) {
                        fprintf(stderr, "malloc madvise failed: ret:
%d, reason: %s\n", ret, strerror(errno));
                }
                usleep(500);
        }
        printf("Done....press ^C\n");
        pause();
        return 0;
}
Note: madvise() system call is not failing here, as mentioned earlier.
I guess the page is aligned with getpagesize().
Then I do this to invoke ksm:
# echo 1 > /sys/kernel/mm/ksm/run
# ./malloc-test.out &
# sleep 5
# echo 0 > /sys/kernel/mm/ksm/run
#

Also, the anon pages in the system shows like this:
BEFORE:
-------------
~ # cat /proc/meminfo | grep -i anon
Active(anon):      40740 kB
Inactive(anon):        0 kB
AnonPages:         40760 kB
AnonHugePages:         0 kB

AFTER MERGING:
--------------------------
~ # cat /proc/meminfo | grep -i anon
Active(anon):        440 kB
Inactive(anon):        0 kB
AnonPages:           188 kB
AnonHugePages:         0 kB

I want to understand the KSM output w.r.t to the above program, and
cross-check if the output is correct.
Can someone help me to understand it?

As of now, what I understood is that:
- I am allocating around 400KB of memory 100 times. That is: 100 * 100
* 4K = 10000 pages (which are all with similar content).
- Output says: 10097 page_sharing happened.
- Pages currently shared is: 3
- So total pages are: 10097 + 3 = 10100

I could not understand from where the additional 100 pages came from?
Also, why some pages are shown as: pages_unshared ?
What can I interpret from this?
And, what does it mean by: pages_volatile:1 ?

Basically, I wanted to understand, is there any problem with the above
output, or it is fine.
If it is fine, how to prove it?



Thanks,
Pintu

On Mon, Sep 17, 2018 at 8:29 PM Mike Rapoport <rppt@...ux.vnet.ibm.com> wrote:
>
> On Mon, Sep 17, 2018 at 05:25:27PM +0530, Pintu Kumar wrote:
> > On Mon, Sep 17, 2018 at 11:46 AM Pintu Kumar <pintu.ping@...il.com> wrote:
> > > > > But still no effect.
> > > > > And I checked LTP test cases. It almost doing the same thing.
> > > > >
> > > > > I observed that [ksmd] thread is not waking up at all.
> > > > > I gave some print inside it, but I could never saw that prints coming.
> > > > > I could not find it running either in top command during the operation.
> > > > > Is there anything needs to be done, to wakw up ksmd?
> > > > > I already set: echo 1 > /sys/kernel/mm/ksm.
> > > >
> > > > It should be echo 1 > /sys/kernel/mm/ksm/run
> > > >
> > >
> > > Oh yes, sorry for the typo.
> > > I tried the same, but still ksm is not getting invoked.
> > > Could someone confirm if KSM was working in 4.9 kernel?
> > >
> >
> > Ok, it's working now. I have to explicitly stop the ksm thread to see
> > the statistics.
> > Also there was some internal patch that was setting vm_flags to
> > VM_MERGABLE thus causing ksm_advise call to return.
> >
> > # echo 1 > /sys/kernel/mm/ksm/run
> > # ./malloc-test.out &
> > # echo 0 > /sys/kernel/mm/ksm/run
> >
> > ~ # grep -H '' /sys/kernel/mm/ksm/*
> > /sys/kernel/mm/ksm/full_scans:105
> > /sys/kernel/mm/ksm/pages_shared:1
> > /sys/kernel/mm/ksm/pages_sharing:999
> > /sys/kernel/mm/ksm/pages_to_scan:100
> > /sys/kernel/mm/ksm/pages_unshared:0
> > /sys/kernel/mm/ksm/pages_volatile:0
> > /sys/kernel/mm/ksm/run:0
> > /sys/kernel/mm/ksm/sleep_millisecs:20
> >
> >
> > However, I have one doubt.
> > Is the above data correct, for the below program?
>
> You have 1 shared page and 999 additional references to that page
>
> > int main(int argc, char *argv[])
> > {
> >         int i, n, size, ret;
> >         char *buffer;
> >         void *addr;
> >
> >         n = 10;
> >         size = 100 * getpagesize();
> >         for (i = 0; i < n; i++) {
> >                 buffer = (char *)malloc(size);
> >                 memset(buffer, 0xff, size);
> >                 madvise(buffer, size, MADV_MERGEABLE);o
>
> This madvise() call should fail because buffer won't be page aligned
>
> >                 addr =  mmap(NULL, size,
> >                            PROT_READ | PROT_EXEC | PROT_WRITE,
> > MAP_PRIVATE | MAP_ANONYMOUS,
> >                            -1, 0);
> >                 memset(addr, 0xff, size);
> >                 ret = madvise(addr, size, MADV_MERGEABLE);
> >                 if (ret < 0) {
> >                         fprintf(stderr, "madvise failed: ret: %d,
> > reason: %s\n", ret, strerror(errno));
> >                 }
> >                 usleep(500);
> >         }
> >         printf("Done....press ^C\n");
> >
> >         pause();
> >
> >         return 0;
> > }
> >
> >
> > Thanks,
> > Pintu
> >
>
> --
> Sincerely yours,
> Mike.
>

Powered by blists - more mailing lists