Hi Jiri,
On 29 August 2013 15:52, Jiri Olsa <jo...@redhat.com> wrote:
> On Tue, Aug 27, 2013 at 10:12:08AM -0600, David Ahern wrote:
>> [Added Jiri who added the feature to perf]
>>
>> On 8/27/13 5:58 AM, Jean Pihet wrote:
>> >Hi!
>> >
>> >I am looking for info and help on the libunwind API.
>> >
>> >What I am trying to achieve is to allow the callchain feature on perf
>> >for ARM (v7 to start with, then ARMv8), from the DWARF info found in
>> >the .debug_frame section.
>> >
>> > From the source code in tools/perf I have added the call to
>> >dwarf_find_debug_frame to load and parse the debug info from
>> >.debug_frame. This works correctly, all IP ranges are found from the
>> >debug code (test program and libraries). Then at some point I get an
>> >assertion: 'perf: dwarf/Gparser.c:459: fetch_proc_info: Assertion
>> >`c->pi.unwind_info' failed.' At that point c->pi is NULL. Cf. below
>> >for more info.
>
> I checked the patch and it looks ok, thought I've never touched
> the .debug_frames code paths which are separated to some extend
> in the libunwind code..
>
>> >
>> >It looks like I am not using the libunwind API as it should, so that
>> >the perf code apparently builds a list of IP ranges to resolve but
>> >cannot use it later on to gather statistics on the runtime info.
>> >
>> >Any idea on how to progress? I am also looking at the *_proc_info API
>> >of libunwind.
>
> keep digging ;-) also I dont have that assert exactly
Indeed!
I found the right implementation ;-p In fact I have to call the
following in order:
- check if the .debug_frame section exists for the binary,
- call dwarf_find_debug_frame to load the debug info,
- call dwarf_search_unwind_table to lookup the called function from the IP.
The patches are coming asap.
It works great now!
Here is an excerpt of the output below:
$ ./tools/perf/perf record -g dwarf -- ../../perf-libunwind/test_app/stress_bt
Total count: 3472419716
[ perf record: Woken up 226 times to write data ]
[ perf record: Captured and wrote 56.488 MB perf.data (~2467981 samples) ]
$ ./tools/perf/perf report --sort symbol --call-graph --stdio
...
# Samples: 7K of event 'cpu-clock'
# Event count (approx.): 1764500000
#
# Overhead Symbol
# ........ ..................................
#
0.00% [.] foo_128
|
|--98.29%-- foo_127
| |
| |--99.29%-- foo_126
| | |
| | |--99.12%-- foo_125
| | | |
| | | |--99.40%-- foo_124
| | | | |
| | | | |--99.13%-- foo_123
| | | | | |
| | | | |
|--99.24%-- foo_122
...
| | | --0.60%-- bar
| | | doit
| | | main
| | | __libc_start_main
| | |
| | --0.88%-- bar
| | doit
| | main
| | __libc_start_main
| |
| --0.71%-- bar
| doit
| main
| __libc_start_main
|
|--0.86%-- bar
| doit
| main
| __libc_start_main
--0.86%-- [...]
The problem is that it is slow as hell (couple of minutes to report
the callchain from a simple test app ;-|). I will look at the sampling
rate and the caching next.
> at dwarf/Gparser.c:459 so maybe try latest code
> git://git.sv.gnu.org/libunwind.git
I am using the version with commit ID 3b8254d3 ('2013-05-19 Matt
Fischer Fix unwind info freeing code in DWARF parser'), which is very
recent.
>
> sorry not to help help you much, you might try to ask
> Arun Sharma <asha...@fb.com> who helped me before.
Ok!
Thanks for your reply and your help!
>
> jirka
Cheers,
Jean
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
