9 Notes on selecting processes in proc.mem and proc.num items
Processes modifying their commandline
Some programs use modifying their commandline as a method for displaying their current activity. A user can see the activity by running ps
and top
commands. Examples of such programs include PostgreSQL, Sendmail, Zabbix.
Let’s see an example from Linux. Let’s assume we want to monitor a number of Zabbix agent processes.
ps
command shows processes of interest as
$ ps -fu zabbix
UID PID PPID C STIME TTY TIME CMD
...
zabbix 6318 1 0 12:01 ? 00:00:00 sbin/zabbix_agentd -c /home/zabbix/ZBXNEXT-1078/zabbix_agentd.conf
zabbix 6319 6318 0 12:01 ? 00:00:01 sbin/zabbix_agentd: collector [idle 1 sec]
zabbix 6320 6318 0 12:01 ? 00:00:00 sbin/zabbix_agentd: listener #1 [waiting for connection]
zabbix 6321 6318 0 12:01 ? 00:00:00 sbin/zabbix_agentd: listener #2 [waiting for connection]
zabbix 6322 6318 0 12:01 ? 00:00:00 sbin/zabbix_agentd: listener #3 [waiting for connection]
zabbix 6323 6318 0 12:01 ? 00:00:00 sbin/zabbix_agentd: active checks #1 [idle 1 sec]
...
Selecting processes by name and user does the job:
$ zabbix_get -s localhost -k 'proc.num[zabbix_agentd,zabbix]'
6
Now let’s rename zabbix_agentd
executable to zabbix_agentd_30
and restart it.
ps
now shows
$ ps -fu zabbix
UID PID PPID C STIME TTY TIME CMD
...
zabbix 6715 1 0 12:53 ? 00:00:00 sbin/zabbix_agentd_30 -c /home/zabbix/ZBXNEXT-1078/zabbix_agentd.conf
zabbix 6716 6715 0 12:53 ? 00:00:00 sbin/zabbix_agentd_30: collector [idle 1 sec]
zabbix 6717 6715 0 12:53 ? 00:00:00 sbin/zabbix_agentd_30: listener #1 [waiting for connection]
zabbix 6718 6715 0 12:53 ? 00:00:00 sbin/zabbix_agentd_30: listener #2 [waiting for connection]
zabbix 6719 6715 0 12:53 ? 00:00:00 sbin/zabbix_agentd_30: listener #3 [waiting for connection]
zabbix 6720 6715 0 12:53 ? 00:00:00 sbin/zabbix_agentd_30: active checks #1 [idle 1 sec]
...
Now selecting processes by name and user produces an incorrect result:
$ zabbix_get -s localhost -k 'proc.num[zabbix_agentd_30,zabbix]'
1
Why a simple renaming of executable to a longer name lead to quite different result ?
Zabbix agent starts with checking the process name. /proc/<pid>/status
file is opened and the line Name
is checked. In our case the Name
lines are:
$ grep Name /proc/{6715,6716,6717,6718,6719,6720}/status
/proc/6715/status:Name: zabbix_agentd_3
/proc/6716/status:Name: zabbix_agentd_3
/proc/6717/status:Name: zabbix_agentd_3
/proc/6718/status:Name: zabbix_agentd_3
/proc/6719/status:Name: zabbix_agentd_3
/proc/6720/status:Name: zabbix_agentd_3
The process name in status
file is truncated to 15 characters.
A similar result can be seen with ps
command:
$ ps -u zabbix
PID TTY TIME CMD
...
6715 ? 00:00:00 zabbix_agentd_3
6716 ? 00:00:01 zabbix_agentd_3
6717 ? 00:00:00 zabbix_agentd_3
6718 ? 00:00:00 zabbix_agentd_3
6719 ? 00:00:00 zabbix_agentd_3
6720 ? 00:00:00 zabbix_agentd_3
...
Obviously, that is not equal to our proc.num[]
name
parameter value zabbix_agentd_30
. Having failed to match the process name from status
file the Zabbix agent turns to /proc/<pid>/cmdline
file.
How the agent sees the “cmdline” file can be illustrated with running a command
$ for i in 6715 6716 6717 6718 6719 6720; do cat /proc/$i/cmdline | awk '{gsub(/\x0/,"<NUL>"); print};'; done
sbin/zabbix_agentd_30<NUL>-c<NUL>/home/zabbix/ZBXNEXT-1078/zabbix_agentd.conf<NUL>
sbin/zabbix_agentd_30: collector [idle 1 sec]<NUL><NUL><NUL><NUL><NUL><NUL><NUL><NUL><NUL><NUL><NUL><NUL><NUL>...
sbin/zabbix_agentd_30: listener #1 [waiting for connection]<NUL><NUL><NUL><NUL><NUL><NUL><NUL><NUL><NUL><NUL>...
sbin/zabbix_agentd_30: listener #2 [waiting for connection]<NUL><NUL><NUL><NUL><NUL><NUL><NUL><NUL><NUL><NUL>...
sbin/zabbix_agentd_30: listener #3 [waiting for connection]<NUL><NUL><NUL><NUL><NUL><NUL><NUL><NUL><NUL><NUL>...
sbin/zabbix_agentd_30: active checks #1 [idle 1 sec]<NUL><NUL><NUL><NUL><NUL><NUL><NUL><NUL><NUL><NUL><NUL><NUL>...
/proc/<pid>/cmdline
files in our case contain invisible, non-printable null bytes, used to terminate strings in C language. The null bytes are shown as “<NUL>” in this example.
Zabbix agent checks “cmdline” for the main process and takes a zabbix_agentd_30
, which matches our name
parameter value zabbix_agentd_30
. So, the main process is counted by item proc.num[zabbix_agentd_30,zabbix]
.
When checking the next process, the agent takes zabbix_agentd_30: collector [idle 1 sec]
from the cmdline
file and it does not meet our name
parameter zabbix_agentd_30
. So, only the main process which does not modify its commandline, gets counted. Other agent processes modify their command line and are ignored.
This example shows that the name
parameter cannot be used in proc.mem[]
and proc.num[]
for selecting processes in this case.
For proc.get[]
item, when Zabbix agent checks “cmdline” for the process name, it will only use part of the name starting from the last slash and until the first space or colon sign. Process name received from cmdline file will only be used if its beginning completely matches the shortened process name in the status
file. The algorithm is the same for both process name in the filter and in the JSON output.
Using cmdline
parameter with a proper regular expression produces a correct result:
$ zabbix_get -s localhost -k 'proc.num[,zabbix,,zabbix_agentd_30[ :]]'
6
Be careful when using proc.get[]
, proc.mem[]
and proc.num[]
items for monitoring programs which modify their command lines.
Before putting name
and cmdline
parameters into proc.get[]
, proc.mem[]
and proc.num[]
items, you may want to test the parameters using proc.num[]
item and ps
command.
Linux kernel threads
Threads cannot be selected with cmdline
parameter in proc.get[]
, proc.mem[]
and proc.num[]
items
Let’s take as an example one of kernel threads:
$ ps -ef| grep kthreadd
root 2 0 0 09:33 ? 00:00:00 [kthreadd]
It can be selected with process name
parameter:
$ zabbix_get -s localhost -k 'proc.num[kthreadd,root]'
1
But selection by process cmdline
parameter does not work:
$ zabbix_get -s localhost -k 'proc.num[,root,,kthreadd]'
0
The reason is that Zabbix agent takes the regular expression specified in cmdline
parameter and applies it to contents of process /proc/<pid>/cmdline
. For kernel threads their /proc/<pid>/cmdline
files are empty. So, cmdline
parameter never matches.
Counting of threads in proc.mem[]
and proc.num[]
items
Linux kernel threads are counted by proc.num[]
item but do not report memory in proc.mem[]
item. For example:
$ ps -ef | grep kthreadd
root 2 0 0 09:51 ? 00:00:00 [kthreadd]
$ zabbix_get -s localhost -k 'proc.num[kthreadd]'
1
$ zabbix_get -s localhost -k 'proc.mem[kthreadd]'
ZBX_NOTSUPPORTED: Cannot get amount of "VmSize" memory.
But what happens if there is a user process with the same name as a kernel thread ? Then it could look like this:
$ ps -ef | grep kthreadd
root 2 0 0 09:51 ? 00:00:00 [kthreadd]
zabbix 9611 6133 0 17:58 pts/1 00:00:00 ./kthreadd
$ zabbix_get -s localhost -k 'proc.num[kthreadd]'
2
$ zabbix_get -s localhost -k 'proc.mem[kthreadd]'
4157440
proc.num[]
counted both the kernel thread and the user process. proc.mem[]
reports memory for the user process only and counts the kernel thread memory as if it was 0. This is different from the case above when ZBX_NOTSUPPORTED was reported.
Be careful when using proc.mem[]
and proc.num[]
items if the program name happens to match one of the thread.
Before putting parameters into proc.mem[]
and proc.num[]
items, you may want to test the parameters using proc.num[]
item and ps
command.