Experimental Physics and
Industrial Control System

Michael Abbott <[email protected]> · Tue, 13 May 2008 06:59:23 +0000 (GMT)

On Mon, 12 May 2008, Andrew Johnson wrote:
> Michael Abbott wrote:
> > 1. Exiting with epicsThreadExitMain() appears to be a very bad thing, at
> > least on RHEL;
> 
> The whole point of the routine epicsThreadExitMain() is to arrange for 
> the main thread to complete _without_ shutting down the other threads om 
> the process.

Ah, I misunderstood that -- however, it's a red herring in the current 
case, as something is going horribly wrong on my machine when we do this.  
I need to dig a bit into the pthread specification, but I'm sure I 
remember that _exit() is supposed to kill *everything* stone cold dead 
without ceremony.

> If you don't ask for an interactive shell there's nothing for the main thread
> to do after loading databases so it calls epicsThreadExitMain().  At this
> point the process looks like a zombie, but it has many other threads still
> running (ls /proc/<pid>/task). 

This is very curious.

$ ps ax | grep defunct
31921 ?        Zl     0:00 [softIoc] <defunct>
$ cd /proc/31921
$ ls -F
attr/    cwd@     fd/       mem         root@  statm   wchan
auxv     environ  loginuid  mounts      smaps  status
cmdline  exe@     maps      mountstats  stat   task/
$ ls task
ls: task: No such file or directory
$

Well now. *That's* an interesting surprise.

> If you kill -9 one of these other threads the whole process should shut 
> down and the zombie goes away.

Yeah.  I think you've broken my machine ;)

What exactly does epicsThreadExitMain() do to try and close the top level 
thread without terminating the process?  When I've had to do this in my 
own processes I've found it easiest to just created a semaphore and block 
on it.

> If you want to be able to shut down a softIoc using Control-C, you have to
> start an interactive shell.  

Or ... it should always be possible (this is fundamental to Linux) to hunt 
down the process and send it a sufficiently aggressive signal (kill -9).  
In this case the process has been persuaded into a state where the rules 
seem to be broken.

> See above, there are still other threads running in the process.

Perhaps not, at least not where anybody can see them.

> The _exit() call does not trigger the atexit() callbacks which trigger the
> shutting down of the other threads. 

Well, it's proving difficult to find a definite statement on this, but 
here's the best I've been able to do so far:

http://publib.boulder.ibm.com/infocenter/systems/index.jsp?topic=/com.ibm.aix.basetechref/doc/basetrf1/pthread_create.htm

"The thread terminates in the following cases: ... The entire process is 
terminated due to a call to ... the ... _exit ... subroutine[s]."

I read that to confirm my understanding that _exit() shouldunconditionally 
kill the calling process and all of its threads.  Why this isn't happening 
on my machine remains a mystery.

Experimental Physics and Industrial Control System

Experimental Physics and
Industrial Control System