Bug 1047

I try your test code using a post TAO 1.2 CVS snapshot.  Unfortunately, I wasn't
able to reproduce the error that you're seeing.   Here's what I get (on Linux):

$ ./test 
CORBA::ORB_init() returned OK
$ cat /dev/null | ./test
CORBA::ORB_init() returned OK

Note that ACE's Service Configurator parser was replaced with a brand new
reentrant parser (generated with GNU Bison) in ACE 5.1.19.  Furthermore,
significant changes were made to the scanner (still GNU Flex based) in that ACE
beta as well.

Can you please try to reproduce this problem with TAO 1.2?

It still seems to happen on ACE/TAO 1.2 but is very Solaris specific.
Here is a stack trace.  I think the problem is the fflush(yyin) call which on
Solaris has the side effect of calling llseek() and setting errno if stdin
is a pipe:

(/opt/SUNWspro/bin/../WS5.0/bin/sparcv9/dbx) where
current thread: t@1
  [1] _lseek64(0x0, 0x0, 0x0, 0x1, 0xff3dc9d0, 0xa1), at 0xff29290c
  [2] _fflush_u(0xff2b63ec, 0x0, 0xff30c524, 0x5257, 0xff3dc9d0, 0xff2b65c0), at 
0xff289cf0
  [3] _fflush_u_iops(0x0, 0xff2b63ec, 0xff2b2118, 0xff2b6584, 0x13, 0x0), at 
0xff289c78
=>[4] ace_yywrap(), line 97 in "Svc_Conf.l"
  [5] ace_yylex(ace_yylval = 0xffbef0a4), line 170 in "Svc_Conf.l"
  [6] ace_yyparse(), line 432 in "bison.simple"
  [7] ACE_Service_Config::process_directives_i(), line 385 in 
"Service_Config.cpp"
  [8] ACE_Service_Config::process_directive(directive = 0x4b6a4 ""), line 418 in 
"Service_Config.cpp"
  [9] TAO_Internal::open_services_i(argc = 1, argv = 0x65f48, 
ignore_default_svc_conf_file = 0, skip_service_config_open = 0), line 256 in 
"TAO_Internal.cpp"
  [10] TAO_Internal::open_services(argc = 1, argv = 0xffbefbcc), line 175 in 
"TAO_Internal.cpp"
  [11] CORBA::ORB_init(argc = 1, argv = 0xffbefbcc, orbid = 0xfe9f7191 "", 
_ACE_CORBA_Environment_variable = CLASS), line 1426 in "ORB.cpp"
  [12] CORBA::ORB_init(argc = 1, argv = 0xffbefbcc, orb_name = (nil)), line 1304 
in "ORB.cpp"
  [13] main(0x1, 0xffbefbcc, 0xffbefbd4, 0x62400, 0x0, 0x0), at 0x2986c


Line 97 of Svc_conf.l is the fflush().
int
yywrap (void)
{
  ::fflush (yyin);  <-- right here.
  yytext[0] = '#';
  yyleng = 0;

  return 1;
}

This doesn't happen on HP-UX, perhaps because their fflush() is more careful 
(ditto for Linux).

The following program simulates the behaviour in minature.
#include <stdio.h>
#include <errno.h>

int main( int argc, char* argv[] )
{
        int res;
        int flush_errno;

        errno = 0;

        res = fflush( stdin );
        flush_errno = errno;
        if ( res != 0 )
        {
                perror( "fflush failed on stdin" );
        }
        else
        {
                fprintf( stderr, "fflush() on stdin succeeded\n" );
        }

        if ( flush_errno != 0 )
        {
                fprintf( stderr, "fflush set errno to %d\n", flush_errno );
                errno = flush_errno;
                perror( "that translates to:" );
        }

        return 0;
}

On HP-UX 11:
% ./a.out
fflush() on stdin succeeded
% cat /dev/null | ./a.out
fflush() on stdin succeeded

On Solaris 2.7:
% ./a.out 
fflush() on stdin succeeded
% cat /dev/null | ./a.out 
fflush() on stdin succeeded
fflush set errno to 29
that translates to:: Illegal seek

I think this is what is happening to the service parser--it is accidentally
causing errno to get set.

Trivia note: FreeBSD 3.2 and 4.2 completely refuse:
% ./a.out
fflush failed on stdin: Bad file descriptor
fflush set errno to 9
% cat /dev/null | ./a.out
fflush failed on stdin: Bad file descriptor
fflush set errno to 9
that translates to:: Bad file descriptor

I forgot to add: the bug won't happen if you have an svc.conf file.
On solaris:

% touch svc.conf
% ./a.out 
CORBA::ORB_init() returned OK
% cat /dev/null | ./a.out 
CORBA::ORB_init() returned OK
% rm svc.conf
% ./a.out 
CORBA::ORB_init() returned OK
% cat /dev/null | ./a.out 
(22421|1) Unable to initialize the Service Configurator: Illegal seek
CORBA::ORB_init threw exception IDL:omg.org/CORBA/INITIALIZE:1.0

I've tracked this down.  Summary:
1. fflush() under solaris can accidentically change errno even though
   it returns no error.
2. fflush() can be called with NULL (I did not know this).
3. fflush() is called with NULL when parsing strings in ace/Svc_conf_l.cpp
   around line 1822 or so:

int
ace_yywrap (void)
{
  ::fflush (ace_yyin);
  ace_yytext[0] = '#';
  ace_yyleng = 0;

  return 1;
}

4. In the case of parsing strings, ace_yyin is NULL.
   I have no idea if the parser meant to do this as calling fflush(NULL)
   is supposed to flush all streams in the process.

5. A patch: change the flush call as follows:
   if (ace_yyin) ::fflush (ace_yyin)

Maybe this bug is fixed in a later version of Solaris?  I'm testing on
"5.7 Generic_106541-04 sun4u sparc SUNW,Ultra-2"

Aceepting the bug for tao-support