[lttng-dev] Getting function names with lttng-ust-cyg-profile.so

[Matthew Khouzam]

Hi Alex and Mathieu,

A year ago we discussed this and said iirc: "the addresses are a good
first step, we'll do more later." and "It would be nice to dump the
stabs into the metadata of the ctf trace"

I would still maintain that this seems to be the way to go.

A technical hurdle that is probably trivial to our high level tracing
vanguards would be to handle dlopens and libs that replace other libs.

I would suggest something like this in TSDL:

location {
    address = 0xdeadbeef;
    name = "get_actual_lost_events";
}

location {
    address = 0xf00ff00f;
    name = "get_pretend_lost_events";
    // optional
    timestart = 123456;
    timeend = 123457;
}


The metadata will get big and dynamic though.

On 13-09-09 07:44 PM, Mathieu Desnoyers wrote:
> We might want to investigate doing a side-program that gathers the
> executables on the system, and lookup the symbols from the ELF. We could
> save those in a bin/ subdirectory of a CTF trace. All we need is
> instrumentation of the dynamic linker, and to know the executable names
> associated with PIDs. There is a UST feature request for dynamic linker
> instrumentation.
>
> Thanks,
>
> Mathieu
>
> * Alexandre Montplaisir (alexmonthy at voxpopuli.im) wrote:
>> Hi all,
>>
>> I've recently started playing with liblttng-ust-cyg-profile.so (aka,
>> getting UST events from -finstrument-functions), and I have to say it's
>> pretty nifty! I haven't done any benchmarks, but it's certainly faster
>> than the typical printf() that people use with it...
>>
>> However, in the resulting trace, one only gets the addresses of the
>> functions. I understand how it's relatively "easy" for the seasoned user
>> to use nm or addr2line to get the actual function names, but would it
>> possible - and how hard would it be - to have this information (function
>> names) directly in the trace?
>>
>>
>> I'm trying to leverage this feature in Eclipse TMF to display a call
>> stack for such UST traces. And to be honest, displaying a call stack
>> with only the function addresses is completely useless, we need the
>> function names.
>>
>> We could have the user import a text file (which he can generate with
>> "nm appname > file.txt" for example). But then he needs the original
>> binary, which he might not have. And that binary needs to be compiled
>> with debugging symbols. If the function name information was already in
>> the trace, it would make the user experience much better, and our job
>> much easier! ;)
>>
>>
>> Thoughts?
>>
>> Thanks,
>> Alex
>>
>> _______________________________________________
>> lttng-dev mailing list
>> lttng-dev at lists.lttng.org
>> http://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev