Hi all...
I thing I found a bug in 2.4 (and someone should check if this applies to 2.5
also...). In short, when running on an HT box, create_elf_tables() changes
u_platform pointer (to separate stacks and avoid cache wars), without
copy_to_user'ing again.
Current code:
if (k_platform) {
platform_len = strlen(k_platform) + 1;
u_platform = p - platform_len;
__copy_to_user(u_platform, k_platform, platform_len);
} else
...
#if defined(__i386__) && defined(CONFIG_SMP)
if(smp_num_siblings > 1)
u_platform = u_platform - ((current->pid % 64) << 7);
///// So original u_platform with data is lost !!!
#endif
...
sp = (elf_addr_t *)(~15UL & (unsigned long)(u_platform)); // Set stack
...
if (k_platform) {
sp -= 2;
NEW_AUX_ENT(0, AT_PLATFORM, (elf_addr_t)(unsigned long) u_platform);
///// What the h**l is copied here after changing u_platform ???
}
Patch below (use an independent stack_top). Boots and works on my dual Xeon,
and glibc behaves correctly:
--- linux/fs/binfmt_elf.c.orig 2002-12-28 00:12:32.000000000 +0100
+++ linux/fs/binfmt_elf.c 2002-12-28 00:32:37.000000000 +0100
@@ -116,11 +116,14 @@
elf_caddr_t *argv;
elf_caddr_t *envp;
elf_addr_t *sp, *csp;
+ char *stack_top;
char *k_platform, *u_platform;
long hwcap;
size_t platform_len = 0;
size_t len;
+ stack_top = p;
+
/*
* Get hold of platform and hardware capabilities masks for
* the machine we are running on. In some cases (Sparc),
@@ -135,8 +138,8 @@
platform_len = strlen(k_platform) + 1;
u_platform = p - platform_len;
__copy_to_user(u_platform, k_platform, platform_len);
- } else
- u_platform = p;
+ stack_top = u_platform;
+ }
#if defined(__i386__) && defined(CONFIG_SMP)
/*
@@ -149,15 +152,14 @@
* processors. This keeps Mr Marcelo Person happier but should be
* removed for 2.5
*/
-
if(smp_num_siblings > 1)
- u_platform = u_platform - ((current->pid % 64) << 7);
+ stack_top -= ((current->pid % 64) << 7);
#endif
/*
* Force 16 byte _final_ alignment here for generality.
*/
- sp = (elf_addr_t *)(~15UL & (unsigned long)(u_platform));
+ sp = (elf_addr_t *)(~15UL & (unsigned long)(stack_top));
csp = sp;
csp -= (1+DLINFO_ITEMS)*2 + (k_platform ? 2 : 0);
#ifdef DLINFO_ARCH_ITEMS
One question: why the 64 value ? Because linux supports 64 processors ?
Why not just 2, because you have two siblings at most ?
TIA
--
J.A. Magallon <[email protected]> \ Software is like sex:
werewolf.able.es \ It's better when it's free
Mandrake Linux release 9.1 (Cooker) for i586
Linux 2.4.21-pre2-jam1 (gcc 3.2.1 (Mandrake Linux 9.1 3.2.1-2mdk))
On 2002.12.28 J.A. Magallon wrote:
> Hi all...
>
[...]
> #if defined(__i386__) && defined(CONFIG_SMP)
> if(smp_num_siblings > 1)
> u_platform = u_platform - ((current->pid % 64) << 7);
> ///// So original u_platform with data is lost !!!
> #endif
> ...
[...]
>
> One question: why the 64 value ? Because linux supports 64 processors ?
> Why not just 2, because you have two siblings at most ?
>
Better, shouldn't this be NR_CPUS ? So you save half stack space on ia32,
or even more with CONFIG_NR_CPUS...
--
J.A. Magallon <[email protected]> \ Software is like sex:
werewolf.able.es \ It's better when it's free
Mandrake Linux release 9.1 (Cooker) for i586
Linux 2.4.21-pre2-jam1 (gcc 3.2.1 (Mandrake Linux 9.1 3.2.1-2mdk))