busybox segfault in __uClibc_main

[Joakim Tjernlund]

> -----Original Message-----
> From: uclibc-bounces at uclibc.org 
> [mailto:uclibc-bounces at uclibc.org] On Behalf Of Linh Dang
> Sent: den 7 december 2005 15:52
> To: uclibc at uclibc.org
> Subject: Re: busybox segfault in __uClibc_main
> 
> Rob Landley <rob at landley.net> wrote:
> 
> > On Tuesday 06 December 2005 11:32, Joakim Tjernlund wrote:
> >> Maybe(need to check my assembler skills I guess), I read it as:
> >> if(r4)
> >> r8 = 0;
> >>
> >> r4 holds argc and I don't get how it can be anything but 
> non zero so 
> >> this should always clear r8 and you don't want that for 
> dynamic apps.
> >
> > Modifying argc and argv the way the current stable busybox does 
> > sucketh mightily.  (It might be a bug that uclibc sets things up so 
> > that doing this segfaults, but it's still a very dodgy 
> thing busybox 
> > is doing.)
> >
> > Does moving to -pre1 fix it for you?  In about three weeks, this 
> > should be the new busybox -stable and it would be nice if 
> we work on 
> > uClibc 0.9.28 on PPC with the 2.6.14 kernel...
> >
> >> Jocke
> >
> > Rob
> 
> I think Joakim does have a point (thanx Joakim) and the main problem
> is:
> 
> - with glibc, early code in _start can distinguish whether it was
>   called from ldso or from the kernel. this is NOT the case with
>   uClibc (atleast with 0.9.28).

Yes, I have seen this but glibc runs on sevral OS:es and this never made
any sense to me.

> 
> - ppc uClibc assumes that all registers are cleared (except for r1)
>   once it reach _start. I'm not sure if it's the case with
>   CONFIG_PREEMPT kernel.  it passed the value of r7 as rtld_fini
>   parameter to __uClibc_main() assuming that r7 will be 0 if it was a
>   staticly linked prog.

uClibc assumes that r7 is cleared and r1 points to argc on stack.
Last time I looked in the kernel it seemed this was the case, but I not sure.
Maybe Linux behaves diffrently for static vs. dynamic apps?
I would really like to see a spec for startup behaviour for linux, I never
found one.

if r1 really pointed to a null word, then the rest of crt is wrong too
and your patch would not work since argc, argv etc. are off by one.
Maybe you should try without CONFIG_PREEMPT?

Also try a simple static hello world app to make trouble shooting easier.

> 
> - so in __uClibc_main(), its rtld_fini is _dl_fini in the case of a
>   dynamicly linked prog, garbage (somehow it point to _start) in a
>   staticly linked prog. That why any program would segfault int exit()
>   when it try to call __rtld_fini.
> 
> - r7 is somehow cleared when a program is ptraced, that why I couldn't
>   debug this problem with gdb nor it would crash under strace.
> 
> 
> Regards
> 
> --
> Linh Dang