From: pavenis AT lanet DOT lv
To: sandmann AT clio DOT rice DOT edu (Charles Sandmann), djgpp-workers AT delorie DOT com
Date: Thu, 9 Aug 2001 18:17:47 +0300
MIME-Version: 1.0
Content-type: text/plain; charset=US-ASCII
Content-transfer-encoding: 7BIT
Subject: Re: Selector Exhaustion
Message-ID: <3B72D3CB.11100.FB5F53@localhost>
In-reply-to: <10108091447.AA13261@clio.rice.edu>
References: <Pine DOT SUN DOT 3 DOT 91 DOT 1010809091937 DOT 7420J-100000 AT is> from "Eli Zaretskii" at Aug 09, 2001 09:23:49 AM
X-mailer: Pegasus Mail for Win32 (v3.12c)
Reply-To: djgpp-workers AT delorie DOT com
Errors-To: nobody AT delorie DOT com
X-Mailing-List: djgpp-workers AT delorie DOT com
X-Unsubscribes-To: listserv AT delorie DOT com
Precedence: bulk

On 9 Aug 2001, at 9:47, Charles Sandmann wrote:

> > It also assumes all the selectors in between belong to the child program, 
> > and thus are not used anymore.  Isn't that a dangerous assumption?
> 
> I worried about it.  Then I did lots of testing and each LDT appeared to
> be independent, even on W95.  So it would only be a problem if we 
> fragment the selectors ourselves.

I did some experiments mostly under Win98SE (and some under 
WinNT 4.0+SP6 and also DOSEMU-1.0.2 under Linux):

Rebuilt make using patched CVS version of DJGPP runtime (patched
doexec.c to cleanup leaked descriptors, patch is included below):

After that I can build development version of gcc-3.0.1 under 
Win98SE without problems (earlier I had to restart make bootstrap 
several times). This time I did all stuff beginning from generating 
source archive for DJGPP without quitting bash.

Run test program for testing descriptor leaks under
	Win98  - works Ok, no more leaks
	WinNT - no improvement 
	DOSEMU - no leaks

I also noticed one strange thing with WinNT 4.0+SP6. If I simply 
allocated descriptor before and after spawn() and cleaned interval
if the difference of both descriptors were correct then it looked that
descriptor leaking was not so big. After using an a similar way as I did
in patch (below) it seems no more so.

Andris

*** djgpp/src/libc/dos/process/dosexec.c~1	Mon Jul 30 14:35:52 2001
--- djgpp/src/libc/dos/process/dosexec.c	Thu Aug  9 13:29:06 2001
*************** static int	     tbuf_selector;
*** 59,64 ****
--- 59,94 ----
  
  static int script_exec(const char *, char **, char **);
  
+ /* Attempt to workaround descriptor leak problem on Win9X  */
+ 
+ static char __desc_map[8192];
+ 
+ static void __build_free_descriptors_map (void);
+ static void __release_extra_descriptors (void);
+ 
+ static void __build_free_descriptors_map (void)
+ {
+      int  i;
+      char * map = __desc_map;
+      for (i=0x0000007; i<0x00010000; i+=8)
+         *map++ = (__dpmi_get_descriptor_access_rights(i) & 0x80) ? 0 : 1;
+ }
+ 
+ static void __release_extra_descriptors (void)
+ {
+      int  i;
+      char * map = __desc_map;
+      for (i=0x0000007; i<0x00010000; i+=8)
+      {
+         if (*map++)
+         {
+             if (__dpmi_get_descriptor_access_rights(i) & 0x80)
+                 __dpmi_free_ldt_descriptor (i);
+ 	}
+      }
+ }
+ 
+ 
  /* Allocate AMT bytes off the transfer buffer.  */
  static unsigned long talloc(size_t amt)
  {
*************** direct_exec_tail(const char *program, co
*** 373,384 ****
--- 403,416 ----
    parm.fcb2_off = fcb2_la & 15;
    dosmemput(&parm, sizeof(parm), parm_la);
  
+   __build_free_descriptors_map ();
    r.x.ax = 0x4b00;
    r.x.ds = program_la / 16;
    r.x.dx = program_la & 15;
    r.x.es = parm_la / 16;
    r.x.bx = parm_la & 15;
    __dpmi_int(0x21, &r);
+   __release_extra_descriptors ();
  #if 0
    if (tbuf_selector)
      __dpmi_free_dos_memory (tbuf_selector);