[RFC] mm: why cat /proc/pid/smaps | grep Rss is different from cat /proc/pid/statm?

[RFC] mm: why cat /proc/pid/smaps | grep Rss is different from cat /proc/pid/statm?

[Xishi Qiu]

[root@localhost c_test]# cat /proc/3948/smaps | grep Rss
Rss:                   4 kB
Rss:                   4 kB
Rss:                   4 kB
Rss:                 796 kB
Rss:                   0 kB
Rss:                  16 kB
Rss:                   8 kB
Rss:                  12 kB
Rss:                 132 kB
Rss:                  12 kB
Rss:                   4 kB
Rss:                   4 kB
Rss:                   4 kB
Rss:                   4 kB
Rss:                  12 kB
Rss:                   0 kB
Rss:                   4 kB
Rss:                   0 kB
[root@localhost c_test]# cat /proc/3948/statm
1042 173 154 1 0 48 0

173 means Rss is 173*4kb=692kb, right?
so why it is different from the sum(1020kb) of "cat /proc/pid/smaps | grep Rss"?

my test code is
...
int main()
{
	sleep(1000);
	return 0;
}

the kernel version is v4.1

Thanks,
Xishi Qiu

Re: [RFC] mm: why cat /proc/pid/smaps | grep Rss is different from cat /proc/pid/statm?

[Xishi Qiu]

On 2016/3/22 22:47, Shakeel Butt wrote:

> 
> On Tue, Mar 22, 2016 at 6:55 AM, Xishi Qiu <qiuxishi@huawei.com <mailto:qiuxishi@huawei.com>> wrote:
> 
>     [root@localhost c_test]# cat /proc/3948/smaps | grep Rss
> 
> The /proc/[pid]/smaps read triggers the traversal of all of process's vmas and then page tables and accumulate RSS on each present page table entry.
> 
>     [root@localhost c_test]# cat /proc/3948/statm
>     1042 173 154 1 0 48 0
> 
> The files /proc/[pid]/statm and /proc/[pid]/status uses the counters (MM_ANONPAGES & MM_FILEPAGES) in mm_struct to report RSS of a process. These counters are modified on page table modifications. However the kernel implements an optimization where each thread keeps a local copy of these counters in its task_struct. These local counter are accumulated in the shared counter of mm_struct after some number of page faults (I think 32) faced by the thread and thus there will be mismatch with smaps file.
> 
> Shakeel

Hi Shakeel,

I malloc and memset 10M, then sleep. It seems that the problem is still exist,
the kernel version is v4.1

[root@localhost c_test]# cat /proc/13746/statm
3603 2767 250 1 0 2609 0
[root@localhost c_test]# cat /proc/13746/smaps | grep Rss
Rss:                   4 kB
Rss:                   4 kB
Rss:                   4 kB
Rss:               10244 kB
Rss:                 924 kB
Rss:                   0 kB
Rss:                  16 kB
Rss:                   8 kB
Rss:                  12 kB
Rss:                 132 kB
Rss:                  12 kB
Rss:                   4 kB
Rss:                   4 kB
Rss:                   4 kB
Rss:                   4 kB
Rss:                   8 kB
Rss:                   0 kB
Rss:                   4 kB
Rss:                   0 kB

Re: [RFC] mm: why cat /proc/pid/smaps | grep Rss is different from cat /proc/pid/statm?

[Xishi Qiu]

On 2016/3/30 15:16, Xishi Qiu wrote:

> On 2016/3/22 22:47, Shakeel Butt wrote:
> 
>>
>> On Tue, Mar 22, 2016 at 6:55 AM, Xishi Qiu <qiuxishi@huawei.com <mailto:qiuxishi@huawei.com>> wrote:
>>
>>     [root@localhost c_test]# cat /proc/3948/smaps | grep Rss
>>
>> The /proc/[pid]/smaps read triggers the traversal of all of process's vmas and then page tables and accumulate RSS on each present page table entry.
>>
>>     [root@localhost c_test]# cat /proc/3948/statm
>>     1042 173 154 1 0 48 0
>>
>> The files /proc/[pid]/statm and /proc/[pid]/status uses the counters (MM_ANONPAGES & MM_FILEPAGES) in mm_struct to report RSS of a process. These counters are modified on page table modifications. However the kernel implements an optimization where each thread keeps a local copy of these counters in its task_struct. These local counter are accumulated in the shared counter of mm_struct after some number of page faults (I think 32) faced by the thread and thus there will be mismatch with smaps file.
>>
>> Shakeel
> 

Hi Shakeel,

I misunderstand your meaning before. I know the reason now.

Thanks,
Xishi Qiu 

> Hi Shakeel,
> 
> I malloc and memset 10M, then sleep. It seems that the problem is still exist,
> the kernel version is v4.1
>