Re: Non-static variables and nested function pointers [bug #28392]

["Grégoire Sutre" <gregoire.sutre@labri.fr>]

[-- Attachment #1: Type: text/plain, Size: 2576 bytes --]

Robert Millan wrote:
> 
> The following snippet (kern/misc.c) comes to mind:
> 
> #ifdef NEED_ENABLE_EXECUTE_STACK
> /* Some gcc versions generate a call to this function
>    in trampolines for nested functions.  */
> void __enable_execute_stack (void *addr __attribute__ ((unused)))
> {
> }
> #endif

I was away from my email, sorry for the delay.  I hunted this
segfault down in the meantime, and I came to the same
conclusion.  If I remove `grub_CHECK_ENABLE_EXECUTE_STACK' from
configure.ac then the segfault in grub-probe disappear.


> We added this for NetBSD in fact.  In that platform, GCC generates references
> to this function, which are usually satisfied by libc, but we don't link with
> libc, so we made it happy with an empty stub.

Yes, I realized that this was NetBSD specific from the archives of the 
mailing list -- kind of ironic ;-)

On my NetBSD/i386 5.0 system, if I compile the attached example, the 
symbol __enable_execute_stack does indeed appear in the executable 
(according to nm).  This holds for gcc 4.1.3 (the system's default 
compiler) and for gcc 4.4.1.  And the symbol __enable_execute_stack is 
found in libgcc_s.so.1.


> But this is only supposed to happen when building real GRUB.  For util/
> stuff, we should use the libc facility instead.  Maybe that's not the case?

I guess that it's not the case: util programs that are linked with 
kern/misc.c do not use the libc's __enable_execute_stack (if 
NEED_ENABLE_EXECUTE_STACK is defined in config.h).


In case that helps, and to continue the discussion regarding
executable stack on NetBSD, I made a few experiments on i386 and
amd64 on a simplified example (see below).  From what I could
read, NetBSD implements memory protection with PaX.

* NetBSD/i386:

                  | with dummy __enable_execute_stack |
                  |       no        |       yes       |
-----------------+-----------------+-----------------+
PaX MPROTECT off |       ok        |    segfault     |
-----------------+-----------------+-----------------+
PaX MPROTECT on  |    segfault     |    segfault     |
-----------------+-----------------+-----------------+

However, in all four cases, there is no segfault if the variable
j in main is declared static.  And in that case, the generated 
executable does not contain the symbol __enable_execute_stack (according 
to nm).

* NetBSD/amd64: in all four cases, no segfault.

Also, on NetBSD/amd64, the segfault problem I experienced in grub-probe 
completely disappears.

Grégoire

[-- Attachment #2: nested.c --]
[-- Type: text/x-csrc, Size: 513 bytes --]

#include <stdio.h>

#ifdef BAD
/* Some gcc versions generate a call to this function
   in trampolines for nested functions.  */
void __enable_execute_stack (void *addr __attribute__ ((unused)))
{
}
#endif

int apply(void (*hook) (int *))
{
   int a = 0;
   hook(&a);
   hook(&a);
   return a;
}

int main (int argc, char *argv[])
{
   /* static */ int j = 5;
   int res;

   void hook(int *n)
   {
      *n = *n + j;
      j--;
   }

   res = apply(hook);
   printf("result: %d, j=%d\n", res, j);
   return 0;
}