Hi,
I already posted this question 2 times on this mailing list under the
subject 'Help on __put_user_asm' but seems like the mailing software
doesn't like such subjects (or the person behind it:).
So here it goes: when I patch the macro in include/asm-i386/uaccess.h
named __put_user_asm using the patch below I cannot 'strace ls' anymore.
The ld.so runtime linker just stops when preparing the program displaying
a nice message; the same message as '/lib/ld-linux.so.2' generates when
ran without any parameters.
The kernel messages of __put_user_asm go to: 0xbffffb44; then the program
stops on the ld.so message.
But, 'strace'-ing of a static compiled program does the job as expected.
What am I doing wrong?
Thanks in advance,
Frank
--- linux-2.4.17/include/asm-i386/uaccess.h Sat Dec 22 09:35:17 2001
+++ linux/include/asm-i386/uaccess.h Mon Dec 24 12:43:22 2001
@@ -194,6 +194,10 @@
* aliasing issues.
*/
#define __put_user_asm(x, addr, err, itype, rtype, ltype) \
+do { \
+ if (current->ptrace & PT_PTRACED) { \
+ printk(KERN_DEBUG "__put_user_asm: %#x\n", (unsigned long)(addr)); \
+ } \
__asm__ __volatile__( \
"1: mov"itype" %"rtype"1,%2\n" \
"2:\n" \
@@ -206,7 +210,8 @@
" .long 1b,3b\n" \
".previous" \
: "=r"(err) \
- : ltype (x), "m"(__m(addr)), "i"(-EFAULT), "0"(err))
+ : ltype (x), "m"(__m(addr)), "i"(-EFAULT), "0"(err)); \
+} while (0)
#define __get_user_nocheck(x,ptr,size) \
Hi,
> > So here it goes: when I patch the macro in include/asm-i386/uaccess.h
> > named __put_user_asm using the patch below I cannot 'strace ls' anymore.
> > The ld.so runtime linker just stops when preparing the program displaying
> > a nice message; the same message as '/lib/ld-linux.so.2' generates when
> > ran without any parameters.
> > The kernel messages of __put_user_asm go to: 0xbffffb44; then the program
> > stops on the ld.so message.
> > But, 'strace'-ing of a static compiled program does the job as expected.
> > What am I doing wrong?
>
> > --- linux-2.4.17/include/asm-i386/uaccess.h Sat Dec 22 09:35:17 2001
> > +++ linux/include/asm-i386/uaccess.h Mon Dec 24 12:43:22 2001
> > @@ -194,6 +194,10 @@
> > * aliasing issues.
> > */
> > #define __put_user_asm(x, addr, err, itype, rtype, ltype) \
> > +do { \
> > + if (current->ptrace & PT_PTRACED) { \
> > + printk(KERN_DEBUG "__put_user_asm: %#x\n", (unsigned long)(addr)); \
> > + } \
> > __asm__ __volatile__( \
>
> This must be clobbering some registers and some users of __put_user_asm
> might not expect that... did you try saving/restoring all regs around this
> 'if'?
I tried to save/restore the eax register before/after the printk function
call because this register is the only register on an i386 that a function
may alter without preserving its previous contents (even when the
function returns void).
What is weird is that an __asm__ __volatile__("nop") within the
if-statement works so it's not the if-statement itself that is causing
the problem, also works is a "push %eax\n\tpop %eax" so again it's not
the stack that is causing the problem.
But, when I try to access the (addr) parameter then execve shows the
ld.so message again. This is very odd because all it does is (in most
cases because GCC optimizes a lot) pushing a general register containing
(addr) onto the stack then call the function (I also tried a void func)
and after the call restore esp.
So, any more advice on this?
Thanks in advance,
Frank.
PS: Yes, I'm a thesis student and I really need this :).
Frank Cornelis <[email protected]> on 12/26/2001 04:12:55 PM
To: vda <[email protected]>
cc: [email protected] (bcc: Amol Lad/HSS)
Subject: Re: Weird __put_user_asm behavior
Hi,
> > So here it goes: when I patch the macro in include/asm-i386/uaccess.h
> > named __put_user_asm using the patch below I cannot 'strace ls' anymore.
> > The ld.so runtime linker just stops when preparing the program displaying
> > a nice message; the same message as '/lib/ld-linux.so.2' generates when
> > ran without any parameters.
> > The kernel messages of __put_user_asm go to: 0xbffffb44; then the program
> > stops on the ld.so message.
> > But, 'strace'-ing of a static compiled program does the job as expected.
> > What am I doing wrong?
>
> > --- linux-2.4.17/include/asm-i386/uaccess.h Sat Dec 22 09:35:17 2001
> > +++ linux/include/asm-i386/uaccess.h Mon Dec 24 12:43:22 2001
> > @@ -194,6 +194,10 @@
> > * aliasing issues.
> > */
> > #define __put_user_asm(x, addr, err, itype, rtype, ltype) \
> > +do { \
> > + if (current->ptrace & PT_PTRACED) { \
> > + printk(KERN_DEBUG "__put_user_asm: %#x\n", (unsigned
long)(addr)); \
> > + } \
> > __asm__ __volatile__( \
>
> This must be clobbering some registers and some users of __put_user_asm
> might not expect that... did you try saving/restoring all regs around this
> 'if'?
I tried to save/restore the eax register before/after the printk function
call because this register is the only register on an i386 that a function
may alter without preserving its previous contents (even when the
function returns void).
>>>> This concept is wrong. Function is free to use/modify any register it
likes. Its not only eax.
Don't expect the value of a register (ebx/ecx etc) to remain same across AFTER a
function call.
What is weird is that an __asm__ __volatile__("nop") within the
if-statement works so it's not the if-statement itself that is causing
the problem, also works is a "push %eax\n\tpop %eax" so again it's not
the stack that is causing the problem.
But, when I try to access the (addr) parameter then execve shows the
ld.so message again. This is very odd because all it does is (in most
cases because GCC optimizes a lot) pushing a general register containing
(addr) onto the stack then call the function (I also tried a void func)
and after the call restore esp.
So, any more advice on this?
Thanks in advance,
Frank.
PS: Yes, I'm a thesis student and I really need this :).
The old code evaluates addr once, your new code evaluates it twice.
Have you tried an inline function instead of a macro?
linux/fs/binfmt_elf.c contains a few lines that probably break if 'addr' is evaluated twice:
<<<<<<<
argv = (elf_caddr_t *) sp;
if (!ibcs) {
__put_user((elf_addr_t)(unsigned long) envp,--sp);
__put_user((elf_addr_t)(unsigned long) argv,--sp);
}
<<<<<<<<<
--
Manfred
Hey,
> The old code evaluates addr once, your new code evaluates it twice.
> Have you tried an inline function instead of a macro?
> linux/fs/binfmt_elf.c contains a few lines that probably break if 'addr'
> is evaluated twice:
> <<<<<<<
> argv = (elf_caddr_t *) sp;
> if (!ibcs) {
> __put_user((elf_addr_t)(unsigned long) envp,--sp);
> __put_user((elf_addr_t)(unsigned long) argv,--sp);
> }
> <<<<<<<<<
Double evaluation of the (addr) expression seems to be my problem, not
register clobbering as I first thought. Although all the macro's that are
using __put_user_asm also use
__typeof__(*(addr)) *__their_local_addr = (addr);
// further use __their_local_addr ...
one apparently needs to repeat this 'localization' of (addr) in order to
have only one evaluation of (addr) at the end.
I always found the macro paragraph in my first C book -'A Book on C'- a
little too short :).
Anyway, thanks to all for the help.
Frank.