Strange socket problem

[huy]

Hi Jeff,

Thanks for your help. Although I haven't confirmed this, I think you 
just hit my nail on the head. I thought os.system was like a totally 
separate process though, i.e nothing is shared. not the usual fork() 
call within the program.

Regards,

Huy

Jeff Epler wrote:
> When using os.system(), files that are open in the parent are available
> in the child, as you can see here in Linux' listing of the files open by
> the child program:
> 
> [jepler at sofa jepler]$ python -c 'f = open("/tmp/test", "w"); print f.fileno(); import os; os.system("ls -l /proc/self/fd")'
> 3
> total 5
> lrwx------  1 jepler jepler 64 Jun 15 07:25 0 -> /dev/pts/2
> lrwx------  1 jepler jepler 64 Jun 15 07:25 1 -> /dev/pts/2
> lrwx------  1 jepler jepler 64 Jun 15 07:25 2 -> /dev/pts/2
> l-wx------  1 jepler jepler 64 Jun 15 07:25 3 -> /tmp/test
> lr-x------  1 jepler jepler 64 Jun 15 07:25 4 -> /proc/3108/fd
> 
> You may be able to set the FD_CLOEXEC flag on the files that should not
> be passed to children, something like this:
>     old = fcntl.fcntl(fd, fcntl.F_GETFD)
>     fcntl.fcntl(fd, fcntl.F_SETFD, old | fcntl.FD_CLOEXEC)
> Refer to a real Unix reference for more information on what FD_CLOEXEC
> does.
> 
> You may want to replace the use of os.system() with fork + close files
> + exec.  Then "myserver.py" will no longer have the listening socket
> descriptor of your cherrypy server.
> 
> Python 2.4's 'subprocess' module may work better in this respect than
> os.system.  It seems to include support for requesting that fds be
> closed in the child (the close_fds parameter to subprocess.Popen)
> 
> Jeff