Re: [Libunwind-devel] libunwind, perf and DWARF callchains

[Mauro Andreolini]

Il giorno ven, 31/01/2014 alle 09.47 -0800, Arun Sharma ha scritto:
> On Thu, Jan 23, 2014 at 7:07 AM, Mauro Andreolini
> <address@hidden> wrote:
> 
> > To which function would 0x7facc1df12a0 match? It does not even seem to
> > be in the mapped address space of "ls". I am clueless.
> 
> Did you try asking the perf guys who did the integration? It'd be
> useful to compile libunwind with --enable-debug and then get a trace
> around what happened during unwinding of this particular address.
Hi, thanks for your time. Yes, I talked to the perf guys and two
problems were solved by patching perf and the kernel event code:
* the 0x0 numbers;
* the numbers in the printf() stack trace.

One address is not unwinded in the dynamic linker code path:

     5.26%       ls  [kernel.kallsyms]  [k]
lock_release                       
                 |
                 --- lock_release
                     __mem_cgroup_try_charge
                     mem_cgroup_charge_common
                     mem_cgroup_newpage_charge
                     do_wp_page
                     handle_mm_fault
                     __do_page_fault
                     do_page_fault
                     do_async_page_fault
                     async_page_fault
                     __init_cpu_features
                     0x7fd2357470de   <---- HERE
                     _dl_relocate_object
                     dl_main
                     _dl_sysdep_start
                     _dl_start
                     _dl_start_user

> If this is happening only in glibc, there is a good chance that there
> is a piece of hand written assembly and some dwarf CFI that libunwind
> doesn't understand yet.
It might be. Are there any instructions on how to recompile libunwind
with debug support and run some tests?

Thanks.
Mauro