The vmRSS value of /proc/$pid/status seems not accurate? Why?

Question

I wrote a test program to help me understand the output field of different memory utilities such as free,ps,top,/proc/$pid/status,/proc/$pid/smaps,/proc/$pid/statm etc. One question finally rose up and I could not manage to understand:

Question: The vmRss field of /proc/$pid/status is not same with the one calculated out of /proc/$pid/smaps (by adding up all the Rss field).

The former value should be source of the RSS output of "ps" command while the latter one source of "pmap -x" command.

The test program I wrote is to create 20 threads with identical procedure in which each thread call malloc(4*1024) 256 times, resulting in a memory footprint of 1MB per thread, therefore 1MB/thread * 20 thread = 20MB totally.

Based on this program, the output of VmRSS from /proc/$pid/status is

VmRSS:     16468 kB

which align with the output of ps

8941  0.0  0.1 4102600 16468 pts/22  Sl+  10:07   0:00 ./a.out

While the add-up of /proc/8941/smaps and output of pmap -x 8941 is:

$ cat /proc/8941/smaps | grep Rss | awk '{print $2}' | awk '{s+=$1} END {printf "%.0f\n", s}' /dev/stdin
22536
$ pmap -x 8941 | tail -n 1
total kB         4102604   22536   20992

The output of free command convinced me that my program do consume 20+MB memory, therefore the RSS value of "ps" and "/proc/$pid/status" does not make sense to me.

Can anybody explain what is happening? Thanks in advance.

Answer 1

Please refer Linux man-pages project at https://www.kernel.org/doc/man-pages/. They can explain what you are seeing.

As of this writing (release 5.13), proc(5) says VmRSS in /proc/[pid]/status is inaccurate.

VmRSS Resident set size. Note that the value here is the sum of RssAnon, RssFile, and RssShmem. This value is inaccurate; see /proc/[pid]/statm above.

proc(5) says RssAnon and RssFile are inaccurate, so VmRSS must be inaccurate.

For accurate values you should refer /proc/[pid]/smaps (man page says so under /proc/[pid]/statm).

And, ps(1) says

The SIZE and RSS fields don't count some parts of a process including the page tables, kernel stack, struct thread_info, and struct task_struct. This is usually at least 20 KiB of memory that is always resident. SIZE is the virtual size of the process (code+data+stack).