Bug 1311682

The test case is very simple:

$ cat /etc/passwd | head -1 /dev/stdin
head: cannot open ‘/dev/stdin’ for reading: No such device or address

while it should show something like:

$ cat /etc/passwd | head -1 /dev/stdin
root:x:0:0:root:/root:/bin/bash

  The alpha and beta branches of https://github.com/att/ast
do not have the problem, but the master branch has the
problem.

  One quick hack is to "#if 0" the chunk in
src/cmd/ksh93/sh/io.c:

#      if _socketpair_shutdown_mode
#         define pipe(v) ((socketpair(AF_UNIX,SOCK_STREAM,0,v)<0||shutdown((v)[1],SHUT_RD)<0||fchmod((v)[1],S_IWUSR)<0||shutdown((v)[0],SHUT_WR)<0||fchmod((v)[0],S_IRUSR)<0)?(-1):0)
#      else
#         define pipe(v) ((socketpair(AF_UNIX,SOCK_STREAM,0,v)<0||shutdown((v)[1],SHUT_RD)<0||shutdown((v)[0],SHUT_WR)<0)?(-1):0)
#      endif

so that it does not redefine the "pipe" identifier;
or change src/cmd/ksh93/features/poll to fail all
tests, and not define _pipe_socketpair.

  But that causes some test errors, for example:
+	io.sh[254]: <# not working for pipes
+	io.sh[339]: read -n3 from pipe not working
+	io.sh[348]: read -n3 from fifo failed -- expected 'a', got 'abc'
+	io.sh[351]: read -n1 from fifo failed -- expected 'b', got 'd'
and causes file descriptor leaks:
+	subshell.sh[578]: file descriptor 11 not close on exec

  The likely commit that corrects the issue is a 280k lines
commit/merge, and is not trivial. Basically it does a significantly
different way of handling descriptors, mostly related to close-on-exec.

  I tried an adaptation of the core idea, but that is too few code,
and likely would require a quite large patch, not viable for
rhel6 and rhel7. The core idea would be to patch sh_pipe like
this:

"""
--- ksh-20120801/src/cmd/ksh93/sh/io.c.orig	2016-02-24 13:54:09.443932505 -0300
+++ ksh-20120801/src/cmd/ksh93/sh/io.c	2016-02-24 13:59:32.807470766 -0300
@@ -893,8 +893,12 @@ int	sh_pipe(register int pv[])
 		errormsg(SH_DICT,ERROR_system(1),e_pipe);
 	pv[0] = sh_iomovefd(pv[0]);
 	pv[1] = sh_iomovefd(pv[1]);
-	shp->fdstatus[pv[0]] = IONOSEEK|IOREAD;
-	shp->fdstatus[pv[1]] = IONOSEEK|IOWRITE;
+	if(pv[0]>2)
+		fcntl(pv[0],F_SETFD,FD_CLOEXEC);
+	if(pv[1]>2)
+		fcntl(pv[1],F_SETFD,FD_CLOEXEC);
+	shp->fdstatus[pv[0]] = IONOSEEK|IOREAD|IOCLEX;
+	shp->fdstatus[pv[1]] = IONOSEEK|IOWRITE|IOCLEX;
 	sh_subsavefd(pv[0]);
 	sh_subsavefd(pv[1]);
 	return(0);
"""
but that causes even worse problems, for example:
	shcomp-basic.ksh[345]: process substitution of compound commands not working
cat: /dev/fd/4: No such file or directory
	shcomp-basic.ksh[354]: process substitution not working outside for or while loop

  Interestingly, I recently rediscovered a chunk almost identical
in beta and alpha ksh, that almost matches the attachment at
https://bugzilla.redhat.com/show_bug.cgi?id=1259898
[table of file descriptors not updated on F_DUPFD_CLOEXEC]

  My impression is that using unix sockets, instead of real
pipes (like bash at least) is hiding issues with leak of
file descriptors.

  I believe there should be ongoing plan to update to ksh-2014,
so, please as well, check this test condition.