How is it possible that kill -9 for a process on Linux has no effect? - linux

Question

I'm writing a plugin to highlight text strings automatically as you visit a web site. It's like the highlight search results but automatic and for many words; it could be used for people with allergies to make words really stand out, for example, when they browse a food site.
But I have problem. When I try to close an empty, fresh FF window, it somehow blocks the whole process. When I kill the process, all the windows vanish, but the Firefox process stays alive (parent PID is 1, doesn't listen to any signals, has lots of resources open, still eats CPU, but won't budge).
So two questions:
How is it even possible for a process not to listen to kill -9 (neither as user nor as root)?
Is there anything I can do but a reboot?
[EDIT] This is the offending process:
USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND
digulla 16688 4.3 4.2 784476 345464 pts/14 D Mar28 75:02 /opt/firefox-3.0/firefox-bin
Same with ps -ef | grep firefox
UID PID PPID C STIME TTY TIME CMD
digulla 16688 1 4 Mar28 pts/14 01:15:02 /opt/firefox-3.0/firefox-bin
It's the only process left. As you can see, it's not a zombie, it's running! It doesn't listen to kill -9, no matter if I kill by PID or name! If I try to connect with strace, then the strace also hangs and can't be killed. There is no output, either. My guess is that FF hangs in some kernel routine but which?
[EDIT2] Based on feedback by sigjuice:
ps axopid,comm,wchan
can show you in which kernel routine a process hangs. In my case, the offending plugin was the Beagle Indexer (openSUSE 11.1). After disabling the plugin, FF was a quick and happy fox again.

Answer 1

As noted in comments to the OP, a process status (STAT) of D indicates that the process is in an "uninterruptible sleep" state. In real-world terms, this generally means that it's waiting on I/O and can't/won't do anything - including dying - until that I/O operation completes.
Processes in a D state will normally only be there for a fraction of a second before the operation completes and they return to R/S. In my experience, if a process gets stuck in D, it's most often trying to communicate with an unreachable NFS or other remote filesystem, trying to access a failing hard drive, or making use of some piece of hardware by way of a flaky device driver. In such cases, the only way to recover and allow the process to die is to either get the fs/drive/hardware back up and running so the I/O can complete or to give up and reboot the system. In the specific case of NFS, the mount may also eventually time out and return from the I/O operation (with a failure code), but this is dependent on the mount options and it's very common for NFS mounts to be set to wait forever.
This is distinct from a zombie process, which will have a status of Z.

Answer 2

Double-check that the parent-id is really 1. If not, and this is firefox, first try sudo killall -9 firefox-bin. After that, try killing the specific process IDs individually with sudo killall -9 [process-id].
How is it even possible for a process not to listen to kill -9 (neiter as user nor as root)?
If a process has gone <defunct> and then becomes a zombie with a parent of 1, you can't kill it manually; only init can. Zombie processes are already dead and gone - they've lost the ability to be killed as they are no longer processes, only a process table entry and its associated exit code, waiting to be collected. You need to kill the parent, and you can't kill init for obvious reasons.
But see here for more general information. A reboot will kill everything, naturally.

Answer 3

Is it possible, that this process is restarted (for example by init) just at the time you kill it?
You can check this easily. If the PID is the same after kill -9 PID then the process wasn't killed, but if it has changed the process has been restarted.

Answer 4

I lately get trapped into a pitfall of Double Fork and had landed to this page before finally finding my answer. The symptoms are identical even if the problem is not the same:
WYKINWYT :What You Kill Is Not What You Thought
The minimal test code is shown below based on an example for an SNMP Daemon
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>
#include <signal.h>
int main(int argc, char* argv[])
{
//We omit the -f option (do not Fork) to reproduce the problem
char * options[]={"/usr/local/sbin/snmpd",/*"-f","*/-d","--master=agentx", "-Dagentx","--agentXSocket=tcp:localhost:1706", "udp:10161", (char*) NULL};
pid_t pid = fork();
if ( 0 > pid ) return -1;
switch(pid)
{
case 0:
{ //Child launches SNMP daemon
execv(options[0],options);
exit(-2);
break;
}
default:
{
sleep(10); //Simulate "long" activity
kill(pid,SIGTERM);//kill what should be child,
//i.e the SNMP daemon I assume
printf("Signal sent to %d\n",pid);
sleep(10); //Simulate "long" operation before closing
waitpid(pid);
printf("SNMP should be now down\n");
getchar();//Blocking (for observation only)
break;
}
}
printf("Bye!\n");
}
During the first phase the main process (7699) launches the SNMP daemon (7700) but we can see that this one is now Defunct/Zombie. Beside we can see another process (7702) with the options we specified
[nils#localhost ~]$ ps -ef | tail
root 7439 2 0 23:00 ? 00:00:00 [kworker/1:0]
root 7494 2 0 23:03 ? 00:00:00 [kworker/0:1]
root 7544 2 0 23:08 ? 00:00:00 [kworker/0:2]
root 7605 2 0 23:10 ? 00:00:00 [kworker/1:2]
root 7698 729 0 23:11 ? 00:00:00 sleep 60
nils 7699 2832 0 23:11 pts/0 00:00:00 ./main
nils 7700 7699 0 23:11 pts/0 00:00:00 [snmpd] <defunct>
nils 7702 1 0 23:11 ? 00:00:00 /usr/local/sbin/snmpd -Lo -d --master=agentx -Dagentx --agentXSocket=tcp:localhost:1706 udp:10161
nils 7727 3706 0 23:11 pts/1 00:00:00 ps -ef
nils 7728 3706 0 23:11 pts/1 00:00:00 tail
After the 10 sec simulated we will try to kill the only process we know (7700). What we succeed at last with waitpid(). But Process 7702 is still here
[nils#localhost ~]$ ps -ef | tail
root 7431 2 0 23:00 ? 00:00:00 [kworker/u256:1]
root 7439 2 0 23:00 ? 00:00:00 [kworker/1:0]
root 7494 2 0 23:03 ? 00:00:00 [kworker/0:1]
root 7544 2 0 23:08 ? 00:00:00 [kworker/0:2]
root 7605 2 0 23:10 ? 00:00:00 [kworker/1:2]
root 7698 729 0 23:11 ? 00:00:00 sleep 60
nils 7699 2832 0 23:11 pts/0 00:00:00 ./main
nils 7702 1 0 23:11 ? 00:00:00 /usr/local/sbin/snmpd -Lo -d --master=agentx -Dagentx --agentXSocket=tcp:localhost:1706 udp:10161
nils 7751 3706 0 23:12 pts/1 00:00:00 ps -ef
nils 7752 3706 0 23:12 pts/1 00:00:00 tail
After giving a character to the getchar() function our main process terminates but the SNMP daemon with the pid 7002 is still here
[nils#localhost ~]$ ps -ef | tail
postfix 7399 1511 0 22:58 ? 00:00:00 pickup -l -t unix -u
root 7431 2 0 23:00 ? 00:00:00 [kworker/u256:1]
root 7439 2 0 23:00 ? 00:00:00 [kworker/1:0]
root 7494 2 0 23:03 ? 00:00:00 [kworker/0:1]
root 7544 2 0 23:08 ? 00:00:00 [kworker/0:2]
root 7605 2 0 23:10 ? 00:00:00 [kworker/1:2]
root 7698 729 0 23:11 ? 00:00:00 sleep 60
nils 7702 1 0 23:11 ? 00:00:00 /usr/local/sbin/snmpd -Lo -d --master=agentx -Dagentx --agentXSocket=tcp:localhost:1706 udp:10161
nils 7765 3706 0 23:12 pts/1 00:00:00 ps -ef
nils 7766 3706 0 23:12 pts/1 00:00:00 tail
Conclusion
The fact that we ignored the double fork mechanism made us think that the kill action did not succeed. But in fact we simply killed the wrong process !!
By adding the -f option ( Do Not (Double) Fork ) all go as expected

Answer 5

ps -ef | grep firefox;
and you can see 3 process, kill them all.

Answer 6

sudo killall -9 firefox
Should work
EDIT: [PID] changed to firefox

Answer 7

You can also do a pstree and kill the parent. This makes sure that you get the entire offending process tree and not just the leaf.