On Fri, 2014-10-24 at 18:33 -0700, Sukadev Bhattiprolu wrote:
> We have some code in the perf tool[1] that uses following
> sequence of calls to try and determine if the return address
> for a function is in the link register (LR) or on the stack:
> 
> 	dwfl = dwfl_begin(&offline_callbacks);
> 
> 	dwfl_report_offline(dwfl, "" exec_file, -1);
> 
> 	mod = dwfl_addrmodule(dwfl, pc);
> 
> 	dwfl_module_eh_cfi(mod, &bias);
> 
> 	/* etc.. */
> 
> where: exec_file is
> 
> 	/usr/lib64/power8/libc-2.17.so (on RHEL7)
> 	/usr/lib64/libc-2.18.so (on Fedora20)
> 
> and 'pc' is an address in libc, say in vfprintf(), rand() etc.
> 
> This sequence of calls works consistently on Fedora20 and fails
> consistently on RHEL7 - where dwfl_addrmodule() returns -1
> (dwfl_errmsg() says "no error").

That should only happen when dwfl == NULL (in which case there should be
an earlier error), or no module contains the given address.

> The versions of elfutils, glibc, gcc are slightly different
> between the two and I have listed them below [2][3][4].
> 
> Is there a difference in the way libc or debug info is
> organized between F20 and RHEL7 ?

I don't believe so. But clearly these are two completely different glibc
versions.

> The Fedora20 is a Power7 system and RHEL7 is a Power8.
> 
> With some debug statements, I see that in dwfl_addrsegment():
> 
> 	lookup() retruns 1,
> 	dwfl->lookup_module is non-null
> 	dwfl_addrsegment() returns -1.
> 
> I can collect more debug if necessary. 

Could you find out what the module load address is for the Dwfl_Module
you create through dwfl_report_offline? dwfl_module_info will give you
the start and end address range. And what is the pc value you provide to
dwfl_addrmodule in both cases?

Thanks,

Mark