Bug 85

fix chan_close/dead/gc:
        user_detach decides:
        send a close only after a close received or always?
        server: delay close until exit message has been sent?
        client: always send close ?

What's wrong with the following obvious patch? This fixes the bug for me
(without this, our network backups hang forever if the tape fills up).

diff -ur openssh-3.5p1.orig/clientloop.c openssh-3.5p1/clientloop.c
--- openssh-3.5p1.orig/clientloop.c     Wed Sep  4 02:32:13 2002
+++ openssh-3.5p1/clientloop.c  Fri Jan 24 12:16:02 2003
@@ -892,6 +892,7 @@
        signal(SIGINT, signal_handler);
        signal(SIGQUIT, signal_handler);
        signal(SIGTERM, signal_handler);
+       signal(SIGPIPE, signal_handler);
        if (have_pty)
                signal(SIGWINCH, window_change_handler);

i think SIGPIPE can happen in x11-fwding and should not lead to exit(1)

I see this bug has now been open for more than four years, with no sign of a fix -- I guess everyone must have a workaround for this issue. Any suggestions, since I haven't been able to find a reliable one...?

I am running into this as well, and have not found a workaround.

The problem seems to be that protocol 2 lacks a way to "half close" a channel, e.g. close its output while keeping its input open.

Markus, does this seem right?

(In reply to comment #6)
> The problem seems to be that protocol 2 lacks a way to "half close" a
> channel, e.g. close its output while keeping its input open.

yes, there is only an 'EOF' message that means:
  I will not send data over this channel,
but there is no way to say: 
  I'm no longer interested in what you want to send to me.

A fix to this bug has been committed and, barring any problems, will be in OpenSSH 5.1:

> CVSROOT:        /cvs
> Module name:    src
> Changes by:     markus@cvs.openbsd.org  2008/05/09 10:16:06
> 
> Modified files:
>         usr.bin/ssh    : session.c
> 
> Log message:
> re-add the USE_PIPES code and enable it.
> without pipes shutdown-read from the sshd does not trigger
> a SIGPIPE when the forked program does a write.
> 
> ok djm@
> 
> 
> CVSROOT:        /cvs
> Module name:    src
> Changes by:     markus@cvs.openbsd.org  2008/05/09 10:17:51
> 
> Modified files:
>         usr.bin/ssh    : channels.c
> 
> Log message:
> error-fd race: don't enable the error fd in the select bitmask
> for channels with both in- and output closed, since the channel
> will go away before we call select();
> report, lots of debugging help and ok djm@
> 
> 
> CVSROOT:        /cvs
> Module name:    src
> Changes by:     markus@cvs.openbsd.org  2008/05/09 10:21:13
> 
> Modified files:
>         usr.bin/ssh    : channels.h clientloop.c nchan.c serverloop.c
> 
> Log message:
> unbreak
> ssh -2 localhost od /bin/ls | true
> ignoring SIGPIPE by adding a new channel message (EOW) that signals
> the peer that we're not interested in any data it might send.
> fixes bz #85; discussion, debugging and ok djm@

Mass update RESOLVED->CLOSED after release of openssh-5.1