Skip to contentSkip to conversation

x86: should 32 bit processes be PER_LINUX32

PATCH In review Andy Whitcroft 3 participants 4 messages Sep 14, 2010 x86

lore
Share & tools
lore
#0 lore raw

cover x86: should 32 bit processes be PER_LINUX32

While testing an x86 64 bit kernel on 32 bit userspace I noticed that I was unexpectedly seeing x86_64 as the architecture as reported in uname. The kernel seems to have code to detect 32 bit userspace and slam this to i686, triggered off the personality PER_LINUX32. However 32 bit processes do not seem to get this flag. I think this is an oversight. The following patch attempts to fix this up.

-apw

Andy Whitcroft (1):
x86: ensure we correctly advertise 32 bit processes in a 64 bit
kernel

 arch/x86/kernel/process_64.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
#1 lore raw

PATCH 1/1 x86: ensure we correctly advertise 32 bit processes in a 64 bit kernel

A number of generic system calls modify their behaviour for 32 bit processes when executed on a 64 bit kernel. For example the uname family return i686 for 32 bit processes but x86_64 for 64 bit processes. These calls determine which return to use based on the personality, specifically they check for PER_LINUX32. x86 does not currently set this personality type, using an x86 specific thread bit to distinguish.

Ensure we use personality PER_LINUX32 for 32 bit processes.

---
 arch/x86/kernel/process_64.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
arch/x86/kernel/process_64.c +1 -1
index 3d9ea53..3793a6a 100644
@@ -518,7 +518,7 @@ void set_personality_ia32(void)
 
 	/* Make sure to be in 32bit mode */
 	set_thread_flag(TIF_IA32);
-	current->personality |= force_personality32;
+	current->personality |= force_personality32 | PER_LINUX32;
 
 	/* Prepare the first "return" to user space */
 	current_thread_info()->status |= TS_COMPAT;
Signature · 1 line
1.7.0.4
#2 lore raw
In reply to Andy Whitcroft
Quoted 6 lines · Andy Whitcroft

While testing an x86 64 bit kernel on 32 bit userspace I noticed that I was unexpectedly seeing x86_64 as the architecture as reported in uname. The kernel seems to have code to detect 32 bit userspace and slam this to i686, triggered off the personality PER_LINUX32. However 32 bit processes do not seem to get this flag. I think this is an oversight. The following patch attempts to fix this up.

No, it's not an oversight. It's possible to hide it behind PER_LINUX32, but that's a bigger hammer that not all users want.

-hpa

#3 lore raw
Quoted 2 lines · H. Peter Anvin
Quoted 1 line · Andy Whitcroft

While testing an x86 64 bit kernel on 32 bit userspace I noticed that I

[...]

Folks, while we're at it, can we also fix the ADDR_LIMIT_3GB personality? With all current kernels it just does not work, returning EINVAL to the second execve() call. It never worked from within a 32bit process, but that at least can be worked around by introducing a small 64bit helper. Now that helper does not work too.

Or just remove it altogether, to stop confusing people... :)

(We finally moved our legacy application from oracle8 (which only exist in 32bit for linux) to oracle10, so this problem has no relevance for us anymore. Before it were quite problematic...)

Thanks!

/mjt