How to breakdown 'Shmem' RAM usage in Linux

Question

I am debugging an issue with a game that is consuming large amounts of memory, essentially all of 24GB available RAM and 8GB of swap is gone a few minutes after start. The memory is not associated with the process itself (according to RSS in top), but seems to be categorized under Shmem in /proc/meminfo. This is based on recording the values as this process is executing, which you can see in the graph here (raw data here).

Right now I am trying to break down Shmem into something more specific. I have looked at the post here which mentions the following are contained under Shmem:

1. tmpfs memory
2. SysV shared memory
3. POSIX shared memory
4. shared anonymous mappings
5. pages used for GEM

I quickly put together this script:

import glob
import os
import re


import psutil


shmem_users = []
# tmpfs memory
disks = psutil.disk_partitions(all=True)
paths = [d.mountpoint for d in disks if d.fstype == 'tmpfs']
tmpfs_usage = sum(psutil.disk_usage(p).used for p in paths)
shmem_users.append({
    'name': 'tmpfs',
    'size': tmpfs_usage
})

# /dev/shm
dev_shm_size = 0
for p in glob.glob('/dev/shm/*'):
    try:
        sz = os.stat(p).st_size
        dev_shm_size += sz
    except:
        # e.g. permission errors
        pass
shmem_users.append({
    'name': '/dev/shm',
    'size': dev_shm_size
})

# TODO: SysV shared memory

# TODO: shared anonymous mappings

# TODO: GEM allocated pages

with open('/proc/meminfo') as f:
    m = re.search(r'Shmem: *(?P<kbs>\d+)', f.read())
    if not m:
        raise RuntimeError('Could not find Shmem value.')
    total_shmem = int(m.group('kbs')) * 1024

print('Total shmem: {}'.format(total_shmem))
total_shmem_accounted_for = sum(u['size'] for u in shmem_users)
print('Total shmem found: {}'.format(total_shmem_accounted_for))
print('Shmem unknown: {}'.format(total_shmem - total_shmem_accounted_for))

I have the process currently SIGSTOPped part-way through eating my RAM, and it returns:

Total shmem: 8450904064
Total shmem found: 681805504
Shmem unknown: 7769098560

So still 7 GBs unaccounted for, hardly closer to understanding what is using up memory.

I have been looking at kernel code for a few hours but have made very little headway in figuring out 2, 4, and 5. I can provide notes if needed. My question: Are these the only other values that could be categorized as "shared memory", and how can I get their specific values?

If it matters my system is:

• Ubuntu 16.04.2 LTS (with KDE 5.5.5)
• 24 GB RAM (+8 GB of swap configured)
• Intel Core i3-6100
• 500GB SSD
• Sapphire Radeon RX470 4GB (with AMDGPU-PRO driver, v17.10)

Solution

For detailing shmem memory usage (and more), you have got the ipcs command.

From man ipcs

NAME ipcs - show information on IPC facilities

SYNOPSIS ipcs [options]

DESCRIPTION ipcs shows information on the inter-process communication facilities for which the calling process has read access. By default it shows information about all three resources: shared memory segments, message queues, and semaphore arrays.

As an example, I am running it here in an Apache server:

$ipcs

------ Message Queues --------
key        msqid      owner      perms      used-bytes   messages    

------ Shared Memory Segments --------
key        shmid      owner      perms      bytes      nattch     status      
0xf4123d98 20643870   root       600        3294       12                      
0xfb123d9b 20676639   root       600        1242       12                      
0x0a123d9d 20709408   root       600        3294       12                      
0x91123da0 20742177   root       600        1242       12                      
0x75123da2 20774946   root       600        3294       12                      
0xbc123da5 20807715   root       600        1242       12                      
0x3d123da7 20840484   root       600        3294       12                      
0x84123daa 20873253   root       600        1242       12                      
0x47123dac 20906022   root       600        3294       12                      
0x0e123daf 20938791   root       600        1242       12                      
0xa6123db1 20971560   root       600        3294       12                      
0x2d123db4 21004329   root       600        1242       12                      
0x77123db6 21037098   root       600        3294       12                      
0x3e123db9 21069867   root       600        1242       12                      
0x80123dbb 21102636   root       600        3294       12                      
0x87123dbe 21135405   root       600        1242       12                      
0xa4123dc0 21168174   root       600        3294       12                      
0xab123dc3 21200943   root       600        1242       12                      
0xfd123dc5 21233712   root       600        3294       12                      
0x44123dc8 21266481   root       600        1242       12                      
0x96123dca 21299250   root       600        3294       12                      
0x1d123dcd 21332019   root       600        1242       12                      
0xa4123dcf 21364788   root       600        3294       12                      
0xab123dd2 21397557   root       600        1242       12                      
0x7a123dd4 21430326   root       600        3294       12                      
0x01123dd7 21463095   root       600        1242       12                      
0xd9123dd9 21495864   root       600        3294       12                      
0x20123ddc 21528633   root       600        1242       12                      
0x86123dde 21561402   root       600        3294       12                      
0x0d123de1 21594171   root       600        1242       12                      

------ Semaphore Arrays --------
key        semid      owner      perms      nsems     
0x00000000 425984     www-data   600        1         
0x00000000 1441793    www-data   600        1         
0x00000000 1474562    www-data   600        1       

Comments

[user2948306]

I did some more /proc spelunking in my own question here. It gets a little bit further, but I still can't account for over 50% of mine (running GNOME on Wayland). My script does not find GEM buffers if they are not currently mmap()ed, so that might be my problem. unix.stackexchange.com/questions/482795/…

[Chris Hunt]

Thank you. This covers Shmem source #2 (SysV shm). Maybe it's too much to ask for all of this in one question...