Date: Sun, 2 May 1999 11:50:02 +0300 (IDT)
From: Eli Zaretskii <eliz AT is DOT elta DOT co DOT il>
X-Sender: eliz AT is
To: djgpp-workers AT delorie DOT com
cc: Charles Sandmann <sandmann AT clio DOT rice DOT edu>
Subject: __dpmi_int causes trouble in signals
Message-ID: <Pine.SUN.3.91.990502114645.563j-100000@is>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
Reply-To: djgpp-workers AT delorie DOT com


A couple of months ago I wrote here about a strange problem with
signals in v2.02 (those who want to see the discussion should search
the mail-archives for a thread with a subject "Exceptions in v2.02"
circa January 1999).

As a simple example, the test program attached below, when compiled
with v2.02 and aborted with Ctrl-BREAK, prints registers that clearly
show that SS:ESP are on the exception stack (and the traceback EIPs
are therefore not printed).  The same program compiled with v2.01
behaves normally, i.e. the full traceback is printed and SS:ESP are on
the normal application stack.

This weekend I finally found the cause of the different behavior.
(The reason it took such a long time was that using the old versions
for the ``usual suspects'', like dpmiexcp.c, exceptn.S and the startup
code, didn't affect the program behavior; so I actually needed to
replace the library modules one by one to find the offending one...)

The function that causes this problem is __dpmi_int (d0300_z.S in the
library).  I attach below its full code, with the added lines marked
with ";+++" and the changed line marked as ";!!!".

Can somebody please tell what was the reason for this change in
__dpmi_int?  (I cannot find anything pertaining to it in the v2.02
changes' description inside wc202.txi.)

If this change does have a good reason, perhaps other parts of the
library need to be changed as well?  For example, exceptn.S switches
to the locked exception stack if DS != SS; perhaps now it should also
accept the DS alias?

Btw, I'm not sure I understand fully why does this change cause a wrong 
SS to be in effect when the traceback is printed.  So if somebody can 
explain that, please do.

I believe that in more complex programs this change causes some grave
bugs, like GPFs in signal handlers when several different signals are
caught, etc.  So I think we need to correct this ASAP.

Here's a test program to demonstrate the phenomenon.  While it runs,
press Ctrl-BREAK twice to get the traceback.

-------------------------- cut here -------------------------------
#include <signal.h>
#include <stdio.h>

char **__crt0_glob_function (char *arg) { return 0; }
void   __crt0_load_environment_file (char *progname) { }
void   __crt0_setup_arguments (void) { }

void quit(int sig)
{
  signal (SIGINT, SIG_DFL);
  while (1)
    puts ("in handler");
}

int main (void)
{
  signal (SIGINT, quit);
  while (1)
    puts ("in main");
  return 0;
}

--------------------------------------------------------------------

Here's the current source of __dpmi_int:

/* Copyright (C) 1998 DJ Delorie, see COPYING.DJ for details */
/* Copyright (C) 1995 DJ Delorie, see COPYING.DJ for details */
#define USE_EBX
#define USE_EDI
#include "dpmidefs.h"

	.comm	___dpmi_int_ss, 2
	.comm	___dpmi_int_sp, 2
	.comm	___dpmi_int_flags, 2
	.lcomm	dpmi_busy, 1

	.text

	FUNC(___dpmi_int)
	ENTER

	pushl	%es
	pushl	%ss					;+++ Line added
	movw	___djgpp_ds_alias, %ax
	movw	%ax, %es
	movw	%ax, %ss				;+++ Line added

	movl	ARG1, %ebx
	movl	$0, %ecx
	movl	ARG2, %edi
	incb	dpmi_busy
	
	cmpb	$1, dpmi_busy
	jne	L_is_busy
	
	movw	___dpmi_int_flags, %ax
	movw	%ax, 0x20(%edi)	/* r.x.flags */
	movw	___dpmi_int_sp, %ax
	movw	%ax, 0x2e(%edi)	/* r.x.sp */
	movw	___dpmi_int_ss, %ax
	movw	%ax, 0x30(%edi)	/* r.x.ss */
	jmp	L_not_busy
	
L_is_busy:
	movw	$0, 0x20(%edi)	/* r.x.flags */
	movw	$0, 0x2e(%edi)	/* r.x.sp */
	movw	$0, 0x30(%edi)	/* r.x.ss */
	
L_not_busy:
	DPMI(0x0300)
	xorl	%eax,%eax

	LEAVEP(decb dpmi_busy; popl %ss; popl %es)	;!!! "popl %ss;" added