mjhjmtu

Could somebody explain to me the difference between kill and killall? Why doesn't killall see what ps shows?

# ps aux |grep db2
root 1123 0.0 0.8 841300 33956 pts/1 Sl 11:48 0:00 db2wdog 
db2inst1 1125 0.0 3.5 2879496 143616 pts/1 Sl 11:48 0:02 db2sysc 
root 1126 0.0 0.6 579156 27840 pts/1 S 11:48 0:00 db2ckpwd 
root 1127 0.0 0.6 579156 27828 pts/1 S 11:48 0:00 db2ckpwd 
root 1128 0.0 0.6 579156 27828 pts/1 S 11:48 0:00 db2ckpwd 

# killall db2ckpwd
db2ckpwd: no process found

# kill -9 1126
# kill -9 1127
# kill -9 1128

System is SuSe 11.3 (64 bit); kernel 2.6.34-12; procps version 3.2.8; killall from PSmisc 22.7; kill from GNU coreutils 7.1

Is this on Linux?

There are actually a few subtly different versions of the command name that are used by ps, killall, etc.

The two main variants are: 1) the long command name, which is what you get when you run ps u; and 2) the short command name, which is what you get when you run ps without any flags.

Probably the biggest difference happens if your program is a shell script or anything that requires an interpreter, e.g. Python, Java, etc.

Here's a really trivial script that demonstrates the difference. I called it mycat:

#!/bin/sh
cat

After running it, here's the two different types of ps.

Firstly, without u:

$ ps -p 5290
 PID TTY ... CMD
 5290 pts/6 ... mycat

Secondly, with u:

$ ps u 5290
USER PID ... COMMAND
mikel 5290 ... /bin/sh /home/mikel/bin/mycat

Note how the second version starts with /bin/sh?

Now, as far as I can tell, killall actually reads /proc/<pid>/stat, and grabs the second word in between the parens as the command name, so that's really what you need to be specifying when you run killall. Logically, that should be the same as what ps without the u flag says, but it would be a good idea to check.

Things to check:

1. what does cat /proc/<pid>/stat say the command name is?


2. what does ps -e | grep db2 say the command name is?


3. do ps -e | grep db2 and ps au | grep db2 show the same command name?

Notes

If you're using other ps flags too, then you might find it simpler to use ps -o comm to see the short name and ps -o cmd to see the long name.

You also might find pkill a better alternative. In particular, pkill -f tries to match using the full command name, i.e. the command name as printed by ps u or ps -o cmd.

killall tries to match on a process name (but is not really that good at the matching part).

And since "ps | grep" and "ps | grep | kill" does a much better job, someone simplified this and created pgrep and pkill. Read that commands like "ps grep" and "ps kill", since that command first ps then grep and if wanted kills.

I had a similar problem but /proc/<pid>/stat contained the expected string. By using strace I could see that killall also accessed /proc/<pid>/cmdline.

I continued to investigate using gdb to find that in my case it failed on a check of my command to the full command including all args found in /proc/<pid>/cmdline. It seemed like that path of the code triggered due to the filename being longer than 15 chars (which is a hardcoded value in the source of killall). I didn't fully investigate if I could somehow getting it to work with killall.

But as mentioned in other comments here pkill is a better alternative that does not have the same issues.

The source code of pkill can be found here https://github.com/acg/psmisc for the interested.

On Ubuntu 16 systems the /proc/pid/stat will contain the name of the thread (which a program can via pthread_setname_np system call.