[Crash-utility] [PATCH] Use backtrace() instead of __builtin_return_address()

[Dave Anderson]

Bernhard Walle wrote:
> * Dave Anderson [2008-05-13 15:16]:
>> A few questions on this one...
>>
>> Do you know if this works OK on ia64, s390 and s390x?
> 
> I made a test program (attached) and attached the result on all
> architectures. In fact, s390 and s390x produced even 100 % correct
> backtraces when the program was optimised with -O2. All other
> architectures leave out frames in that case, but I think that's
> limitation when frame pointers are omitted that is impossible to work
> around.
> 
>> Compiling with warnings turned on yields:
>>
>>    defs.h:1804: warning: implicit declaration of function ‘backtrace’
> 
> I forgot the <execinfo.h> include.
> 
>> I haven't looked at the glibc sources, but I'm presuming it's
>> a "void backtrace(int)".
> 
> Indeed, it's int backtrace(void **buffer, int size). In my patch I
> ignored the return value. The new patch (attached) corrects that
> mistake. I'm not sure if it's necessary to zero out the rest since the
> memory already was initialised with 0, but to be on the save side I
> think it's not bad to do it.
> 
> Since sizeof(unsigned long) == sizeof(void *) on all architectures
> crash supports, the "wrong" type of buffer should be ok. In fact,
> __builtin_return_address() returns also an address, i.e. void * and
> not a long. Conforming strictly to C99, we should use uintptr_t. I
> don't know what the crash programming guide lines say, i.e. how old the
> compiler can be where crash is used. I think we should not introduce
> a C99 requirement because of that small patch here.
> 
> 
> A few words on the test program: I used backtrace_symbols_fd() to
> print out the backtrace here because I was too lazy to "port" the crash
> nm magic to the test program. However, the program needs to be compiled
> with -rdynamic to make that work. That's also documented in
> backtrace(3) manual page. So for crash, the nm magic is a big "ugly"
> but it works better than the backtrace_symbols_fd() here. But for the
> test program I was only interesting in the values backtrace() produces,
> not in resolving, so I think that should be ok.
> 
> 

Thanks for digging into this.  I agree with you on all counts.

One final question: does the remaining call to __builtin_return_address(0)
in tools.c:getbuf() fail in your configuration as well?

Dave