Path: utzoo!mnetor!uunet!husc6!mit-eddie!ll-xn!ames!amdahl!oliveb!sun!betelgeuse!jamesa
From: jamesa%betelgeuse@Sun.COM (James D. Allen)
Newsgroups: comp.unix.questions
Subject: Re: System crash when using sockets
Message-ID: <37990@sun.uucp>
Date: 6 Jan 88 09:34:51 GMT
References: <86@cui.UUCP>
Sender: news@sun.uucp
Lines: 38
Summary: Don't Exit with un-Accepted Connections

In article <86@cui.UUCP>,  Stephane Spahni writes:
> I have a question about sockets on Unix: I tried to use a socket between
> two processes, and regularly crash the system (i.e. the system is looping
> somewhere in the kernel and is doing nothing else !). The only way to
> recover is to halt the system (^P HALT on Vax, L1-A on Sun) and reboot
> it !

Stephane's program neglects to recognize that read returning 0 (EOF) is
different from errno 35 (WOULDBLOCK).  In the former case, the other end
has closed its connection so it's time to try another accept().

This is no excuse for the kernel hang of course.  (The faulty zomby disables
interrupts; otherwise it would just be another "something is hung (wont die)".)
The fact that this bug has not been fixed yet suggests that very few
programmers out there are trying to "accept connections".

The hang arises when soclose() tries to empty so_q but can't.  I don't have an
authoritative fix but sofree() will eventually do the cleanup so just deleting
the first block of code in soclose() lets my machine run Stephane's "test".

For you kids who want to duplicate Stephane's crash, the following should
suffice:

#include <sys/types.h>
#include <sys/socket.h>

short ugly[] = { AF_UNIX, 'GU', 'YL', 0 };

main()
{
	int sr = socket(AF_UNIX, SOCK_STREAM, 0);
	int sw = socket(AF_UNIX, SOCK_STREAM, 0);

	bind(sr, ugly, 6);
	listen(sr, 1);
	connect(sw, ugly, 6);
}
/* -- james allen  */