Work around Linux kernel bug provoked by nchan.c

[Kevin Steves]

hp-ux 10.20 and 11 seem to have a similar issue in that
shutdown(sock, SHUT_RD) return EINVAL when the write side has been
previously closed.  you can dup this with:

ssh -R 2300:bar:23 foo
telnet 0 2300

Dec 22 22:04:18 jenny sshd[21722]: error: channel
1: chan_shutdown_read: shutdown() failed for fd13 [i1 o128]: Invalid
argument

regarding the patch below, why do we test for this:

+			|| c->ostate == CHAN_OUTPUT_OPEN
+			|| c->ostate == CHAN_OUTPUT_WAIT_DRAIN)) 

vs. just checking for ENOTCONN?

On Sun, 23 Jul 2000, Zack Weinberg wrote:
: The Linux implementation of TCP sockets has a bug which causes
: shutdown(sock, SHUT_RD) to fail spuriously (ENOTCONN) if the write
: side of the socket has already been shut down.  If you are using SSH
: port forwarding to tunnel HTTP through a firewall, nchan.c will tickle
: this bug once for every HTTP exchange.  You will therefore get lots of
: useless, annoying error messages:
: 
: channel 2: chan_shutdown_read: shutdown() failed for fd7 [i1 o128]: Transport endpoint is not connected
: 
: Here's a complete debugging trace of one such forwarded channel.  This
: is SSH1 protocol; I haven't got the setup to do SSH2 yet.  The remote
: server is 2.0.12 F-SECURE SSH on Solaris 2.6; I am using openssh
: 2.1.1p4 with kernel 2.2.16.
: 
: debug: channel 2: new [listen port 3128 for webcache.example.com port 3128, connect from localhost port 1817]
: debug: channel 2: rcvd ieof
: debug: channel 2: output open -> drain
: debug: channel 2: obuf empty
: debug: channel 2: output drain -> closed
: debug: channel 2: send oclose
: debug: channel 2: close_write
: debug: channel 2: read<=0 rfd 7 len 0
: debug: channel 2: read failed
: debug: channel 2: input open -> drain
: debug: channel 2: close_read
: channel 2: chan_shutdown_read: shutdown() failed for fd7 [i1 o128]: Transport endpoint is not connected
: debug: channel 2: input: no drain shortcut
: debug: channel 2: ibuf empty
: debug: channel 2: input drain -> wait_oclose
: debug: channel 2: send ieof
: debug: channel 2: rcvd oclose
: debug: channel 2: input wait_oclose -> closed
: debug: channel 2: full closed
: 
: I'd appreciate it if the appended patch could be applied.  It causes
: ssh to recognize the bug and not emit the error message.
: 
: [I've reported the bug to the kernel developers but they do not seem
: interested in fixing it.]
: 
: zw
: 
: --- openssh-2.1.1p4.orig/nchan.c	Thu Jun 22 04:32:31 2000
: +++ openssh-2.1.1p4/nchan.c	Sun Jul 23 09:42:23 2000
: @@ -483,7 +483,12 @@
:  		return;
:  	debug("channel %d: close_read", c->self);
:  	if (c->sock != -1) {
: -		if (shutdown(c->sock, SHUT_RD) < 0)
: +		/* shutdown(sock, SHUT_READ) may return ENOTCONN if the
: +		   write side has been closed already.  */
: +		if (shutdown(c->sock, SHUT_RD) < 0
: +		    && (errno != ENOTCONN
: +			|| c->ostate == CHAN_OUTPUT_OPEN
: +			|| c->ostate == CHAN_OUTPUT_WAIT_DRAIN))
:  			error("channel %d: chan_shutdown_read: shutdown() failed for fd%d [i%d o%d]: %.100s",
:  			    c->self, c->sock, c->istate, c->ostate, strerror(errno));
:  	} else {
: