Re: Application launching and pasteboard slow down

[Richard Frith-Macdonald]

On Monday, April 15, 2002, at 10:53 PM, Richard Frith-Macdonald wrote:

> On Monday, April 15, 2002, at 08:24 PM, Chris B. Vetter wrote:
>
> > gdomap is launched as
> >
> >   PREFIX=$(expr $0 : "\(/.*\)/etc/rc\.d/$(basename $0)\$");
> >   IFFILE=${PREFIX}/etc/gdomap_if
> >
> >   ${GNUSTEP_SYSTEM_ROOT}/Makefiles/GNUstep.sh
> >
> >   ${GNUSTEP_SYSTEM_ROOT}/Tools/ix86/freebsd/gdomap -a ${IFFILE}
>                                                      ^^
> This is not the flag to run with probing disabled, it's the one
> to specify the local IP address(es) on machines with broken
> networking which doesn't support interrogation of the network
> interfaces.
>
>
> I notice you are running on freebsd. I have heard of a version of
> freebsd where the networking was broken so that you couldn't detect
> that a port was in use by trying to bind to it.  This would cause
> multiple copies of the servers to run, because the O/S would tell
> gdomap that the old copies had died when they hadn't.
>
> The person who had that problem fixed it by changing to another
> version of the operating system where the networking was not broken.
>
> If you want to stick with the broken O/S, you can work-around this
> problem by commenting out the lines setting SO_REUSEADDR in the
> GNUstep networking code.  The only drawback of this is that, if
> a server is shut down, it will take a few minutes before the ip
> code in the kernel will let it be restarted on the same port -
> this is not usually a difficulty.

I have added an autoconf check for the SO_REUSEADDR bug in the base 
library,
and modified the networking code to use the simple workaround of not 
using
SO_REUSEADDR on systems where it is broken (windoze, some FreeBSDs, and 
some
versions of Solaris AFAIK).

Explanation of the bug ... the SO_REUSEADDR option is set on a network 
socket
to say that, if the socket is in a TIME_WAIT state (ie it has recently 
been
closed, but the handshake for the close with the other end of the socket 
has
not been completed) then the socket address/port may be re-used 
immediately.
This permits a service to be restarted on the same address/port without 
having
to wait for the (typically 1 to 4 minute) timeout.

On some systems (windoze, some versions of FreeBSD, some of Solaris, and 
perhaps
a few others), using SO_REUSEADDR actually permits the address/port to 
be re-used
*no-matter-what-state-it's-in*.
This has two effects -
1. a security issue - since other processes can bind to the same address
as your process and steal packets.
2. more importantly for us, we can't attempt to bind to an address/port 
in
order to find out if there is already a process using it - our bind 
attempt
will succeed when it really shouldn't.  This means that you can start 
multiple
copies of servers using the same name ... since the nameserver process 
will
always think that the old server has died.

Not using SO_REUSEADDR  will leave most networking unaffected, but if a
gdomap process is killed, it will not be able to restart until the system
timeout has expired.