On Mon, Oct 06, 2008 at 08:13:19AM +0200, Andi Kleen wrote:
> "Kirill A. Shutemov" <kirill@shutemov.name> writes:
> >
> >> 
> >> but more generally, we already have ADDR_LIMIT_3GB support on x86.
> >
> > Does ADDR_LIMIT_3GB really work?
> 
> As Arjan pointed out it only takes effect on exec()
> 
> andi@basil:~/tsrc> cat tstack2.c
> #include <stdio.h>
> int main(void)
> {
>         void *p = &p;
>         printf("%p\n", &p);
>         return 0;
> }
> andi@basil:~/tsrc> gcc -m32 tstack2.c  -o tstack2
> andi@basil:~/tsrc> ./tstack2 
> 0xff807d70
> andi@basil:~/tsrc> linux32 --3gb ./tstack2 
> 0xbfae2840

Which kernel do you use?
Does it work only when compiled with -m32?

$ cat 1.c
#include <stdio.h>
int main(void)
{
        void *p = &p;
        printf("%p\n", &p);
        return 0;
}
$ gcc 1.c
$ linux32 --3gb ./a.out
0x7fffa667e7b8

> >> Why 
> >> should support for ADDR_LIMIT_32BIT be added?
> >
> > It's useful for user mode qemu when you try emulate 32-bit target on 
> > x86_64. For example, if shmat(2) return addres above 32-bit, target will
> > get SIGSEGV on access to it.
> 
> The traditional way in mmap() to handle this is to give it a search
> hint < 4GB and then free the memory again/fail if the result was >4GB.

mmap() has MAP_32BIT flag on x86_64.

> Unfortunately that doesn't work for shmat() because the address argument
> is not a search hint, but a fixed address. 
> 
> I presume you need this for the qemu syscall emulation. For a standard
> application I would just recommend to use mmap with tmpfs instead
> (sysv shm is kind of obsolete). For shmat() emulation the cleanest way
> would be probably to add a new flag to shmat() that says that address
> is a search hint, not a fixed address. Then implement it the way recommended
> above.

I prefer one handle to switch application to 32-bit address mode. Why is it
wrong?

-- 
Regards,  Kirill A. Shutemov
 + Belarus, Minsk
 + ALT Linux Team, http://www.altlinux.com/