[lttng-dev] [RFC] perf to ctf converter

[Sebastian Andrzej Siewior]

On 07/14/2014 04:15 PM, Jiri Olsa wrote:
>> for more data while reading the "events" traces. The latter will be
>> probably replaced by https://lkml.org/lkml/2014/4/3/217.
>> Babeltrace needs only
>>     "ctf-writer: Add support for the cpu_id field"
>>     https://www.mail-archive.com/lttng-dev@lists.lttng.org/msg06057.html
> 
> any idea when this one will land in babeltrace git tree?

Need to re-do them the way they asked. Could take some time. However I
wanted first to make sure it make sense to continue that approach.

>>
>> for the assignment of the CPU number.
>>
>> The pickle step is nice because I see all type of events before I
>> start writing the CTF trace and can create the necessary objects. On
>> the other hand it eats a lot of memory for huge traces so I will try to
>> replace it with something that saves the data in a streaming like
>> fashion.
>> The other limitation is that babeltrace doesn't seem to work with
>> python2 while perf doesn't compile against python3.
>>
>> What I haven't figured out yet is how to pass to the meta environment
>> informations that is displayed by "perf script --header-only -I" and if
>> that information is really important. Probably an optional python
>> callback will do it.
>>
>> The required steps:
>> |   perf record -e raw_syscalls:* w
>> |   perf script -s ./to-pickle.py
>> |   ./ctf_writer
> 
> I made similar effort in C:
> 
> ---
> I made some *VERY* early perf convert example, mostly to try the ctf-writer
> interface.. you can check in here:
>   https://git.kernel.org/cgit/linux/kernel/git/jolsa/perf.git/log/?h=perf/ctf_2

Let me try it, maybe I can migrate my effort into one code basis.

> It's able to convert single event (HW type) perf.data file into CTF data,
> by adding just one integer field "period" and single stream, like:
> 
>   [jolsa at krava perf]$ LD_LIBRARY_PATH=/opt/libbabeltrace/lib/ ./perf data convert --to-ctf=./ctf-data
>   ...
>   [jolsa at krava babeltrace]$ /opt/libbabeltrace/bin/babeltrace /home/jolsa/kernel.org/linux-perf/tools/perf/ctf-data
>   [08:14:45.814456098] (+?.?????????) cycles: { }, { period = 1 }
>   [08:14:45.814459237] (+0.000003139) cycles: { }, { period = 1 }
>   [08:14:45.814460684] (+0.000001447) cycles: { }, { period = 9 }
>   [08:14:45.814462073] (+0.000001389) cycles: { }, { period = 182 }
>   [08:14:45.814463491] (+0.000001418) cycles: { }, { period = 4263 }
>   [08:14:45.814465874] (+0.000002383) cycles: { }, { period = 97878 }
>   [08:14:45.814506385] (+0.000040511) cycles: { }, { period = 1365965 }
>   [08:14:45.815056528] (+0.000550143) cycles: { }, { period = 2250012 }
> ---
> 
> the goals for me is to have a convert tool, like in above example
> perf data command and support in perf record/report to directl
> write/read ctf data
> 
> Using python for this seems nice.. I'm not experienced python coder,
> so just small comments/questions

python looked nice because I saw libraries / interfaces on both sides.

> SNIP
> 
>> +list_type_h_uint64 = [ "addr" ]
>> +
>> +int32_type = CTFWriter.IntegerFieldDeclaration(32)
>> +int32_type.signed = True
>> +
>> +uint64_type = CTFWriter.IntegerFieldDeclaration(64)
>> +uint64_type.signed = False
>> +
>> +hex_uint64_type = CTFWriter.IntegerFieldDeclaration(64)
>> +hex_uint64_type.signed = False
>> +hex_uint64_type.base = 16
>> +
>> +string_type = CTFWriter.StringFieldDeclaration()
>> +
>> +events = {}
>> +last_cpu = -1
>> +
>> +list_ev_entry_ignore = [ "common_s", "common_ns", "common_cpu" ]
>> +
>> +# First create all possible event class-es
> 
> this first iteration could be handled in the to-pickle step,
> which could gather events description and store/pickle it
> before the trace data

yes.

>> +for entry in trace:
>> +    event_name = entry[0]
>> +    event_record = entry[1]
>> +
>> +    try:
>> +        event_class = events[event_name]
>> +    except:
>> +        event_class = CTFWriter.EventClass(event_name);
>> +        for ev_entry in sorted(event_record):
>> +            if ev_entry in list_ev_entry_ignore:
>> +                continue
>> +            val = event_record[ev_entry]
>> +            if isinstance(val, int):
>> +                if ev_entry in list_type_h_uint64:
>> +                    event_class.add_field(hex_uint64_type, ev_entry)
>> +                else:
>> +                    event_class.add_field(int32_type, ev_entry)
>> +            elif isinstance(val, str):
>> +                event_class.add_field(string_type, ev_entry)
> 
> 
> SNIP
> 
>> +
>> +def process_event(event_fields_dict):
>> +    entry = []
>> +    entry.append(str(event_fields_dict["ev_name"]))
>> +    fields = {}
>> +    fields["common_s"] = event_fields_dict["s"]
>> +    fields["common_ns"] = event_fields_dict["ns"]
>> +    fields["common_comm"] = event_fields_dict["comm"]
>> +    fields["common_pid"] = event_fields_dict["pid"]
>> +    fields["addr"] = event_fields_dict["addr"]
>> +
>> +    dso = ""
>> +    symbol = ""
>> +    try:
>> +        dso = event_fields_dict["dso"]
>> +    except:
>> +        pass
>> +    try:
>> +        symbol = event_fields_dict["symbol"]
>> +    except:
>> +        pass
> 
> I understand this is just a early stage, but we want here
> detection of the all event arguments, right?

Yes. The CTF writer is stupid and takes all arguments as-is and passes
it over the babeltrace part of CTF writer. This works well for the
ftrace events (handled by trace_unhandled()).


> I wonder we could add separated python callback for that

This (the to pickle part) tries come up with the common basis for the
CPU events. Therefore it renames the first few arguments (like s to
common_s) to make it consistent with the ftrace events.
The dso and symbol members look optional depending whether or not this
data was available at trace time. I *think* those may change within a
stream say if one library has debug symbols available and the other
does not. So I have no idea how you plan specific callbacks for those.

> thanks,
> jirka

Sebastian