Sockets and interrupt driven I/O

[Chris Torek]

In article <1011 at m1.cs.man.ac.uk> HoldswoS at r4.cs.man.ac.uk (Sean Holdsworth)
writes:
>... if I write to a socket and get back an EWOULDBLOCK error I have no
>way of knowing, other than polling, when the socket again becomes writable.
>I had hoped that when the state of the socket changed from blocked to
>unblocked that a SIGIO would be generated but from my experiments this
>appears not to be the case.

If you examine the file netinet/tcp_input.c, you will find code of the
form:

	if (act > so->so_snd.sb_cc) {
		tp->snd_wnd -= so->so_snd.sb_cc;
		sbdrop(&so->so_snd, (int)so->so_snd.sb_cc);
		ourfinisacked = 1;
	} else {
		sbdrop(&so->so_snd, acked);
		tp->snd_wnd -= acked;
		ourfinisacked = 0;
	}
-->	if ((so->so_snd.sb_flags & SB_WAIT) || so->so_snd.sb_sel)
		sowwakeup(so);
	tp->snd_una = ti->ti_ack;

The line marked `-->' above is the cuprit:  For the sake of efficiency,
the TCP code calls sowwakeup(so) (a macro for sowakeup(so, &so->so_snd))
only if a process is waiting for data to go out, or is selecting, or was
selecting not long ago.  When a process uses asynchronous I/O, however,
neither SB_WAIT nor sb_sel are set.

There are a number of ways around the problem.  One of them involves no
kernel changes: simply arrange for some process to select on that socket.
Any process will do, including your own.  Thus:

	omask = sigblock(sigmask(SIGIO));
	if ((n = write(socket_fd, buf, count)) < 0 && errno == EWOULDBLOCK) {
		fd_set out;
		struct timeval tv;
		FD_ZERO(&out);
		FD_SET(socket_fd, &out);
		tv.tv_sec = 0;
		tv.tv_usec = 0;
		n = select(socket_fd + 1, (fd_set *)0, &out, (fd_set *)0, &tv);
		if (n > 0) {
			/* can write now, try again: must have unblocked
			   while we were fiddling with select() */
			n = write(socket_fd, buf, count);
			if (n < 0 && errno == EWOULDBLOCK)
				... now what? ...
		}
		/* cannot write, but made so->so_snd.sb_sel non nil so that
		   the tcp code will call sowakeup() later */
		n = 0;	/* clobber the error */
	}
	if (n < 0)
		... handle output error ...
	(void) sigsetmask(omask);

Simpler fixes, if you have kernel source, are to remove the test from
tcp_input.c or to assert SB_WAIT in sosend() when returning EWOULDBLOCK.
(The latter would preserve efficiency, and protect any other code that
has the same bug as tcp_input.c, such as the XNS code [spp_usrreq.c].)
The bug has been fixed in 4.4BSD by removing the test from both
netinet/tcp_input.c and netns/spp_usrreq.c.
-- 
In-Real-Life: Chris Torek, Univ of MD Comp Sci Dept (+1 301 454 7163)
Domain:	chris at cs.umd.edu	Path:	uunet!mimsy!chris