Tracepoints (and python)

Tracepoints (and python)

[Marc Brünink]

I have no idea about gdb internals. I m trying to learn, though. Questions:

1. Is there any technical reason why tracing is only possible on remote targets? Or is is "just" not implemented for anything else.
2. Anyone working on making tracepoints available to the python interface? Any pitfalls to expect?

Marc

Re: Tracepoints (and python)

[Tom Tromey]

>>>>> "Marc" == Marc Brünink <marc@nus.edu.sg> writes:

Marc> 1. Is there any technical reason why tracing is only possible on
Marc> remote targets? Or is is "just" not implemented for anything else.

I think the only reason is that nobody has done it.

Marc> 2. Anyone working on making tracepoints available to the python
Marc> interface? Any pitfalls to expect?

Nobody has worked on this.
I wouldn't predict any particularly tricky problems, at least not for
representing the tracepoints themselves.  I'm less sure about
representing the trace state.

Tom

Re: Tracepoints (and python)

[Hui Zhu]

On Thu, Dec 20, 2012 at 10:32 PM, Tom Tromey <tromey@redhat.com> wrote:
>>>>>> "Marc" == Marc Brünink <marc@nus.edu.sg> writes:
>
> Marc> 1. Is there any technical reason why tracing is only possible on
> Marc> remote targets? Or is is "just" not implemented for anything else.
>
> I think the only reason is that nobody has done it.
>
> Marc> 2. Anyone working on making tracepoints available to the python
> Marc> interface? Any pitfalls to expect?
>
> Nobody has worked on this.
> I wouldn't predict any particularly tricky problems, at least not for
> representing the tracepoints themselves.  I'm less sure about
> representing the trace state.
>

I wrote some python code with tracepoint to parse the traceframe from
target (my target is KGTP).  You can get a example in
http://code.google.com/p/kgtp/wiki/hotcode

I have 3 issues with it:
1. Not sure about upstream, but when I use it, it is hard to setup
tracepoint action with python.  How I handle it is write tracepoint
and action to a GDB source file and call soure command in python
script.

2. Another is the performance issue.  If you have a lot of traceframe
need to be analysed by python,  you will get it.  And actually, the
url that I posted is about a python+gdb performance issue and a
solution of it.

3. The last one is if you can handle traceframe with python, maybe you
think you can keep analyse the traceframe from target.  But current
tracepoint doesn't it.  You must "tstart", "tstop", parse.
What I handle it is let KGTP supply a special interface in tfile
format.  Then GDB can keep get traceframe from it without stop
tracepoint.

> Tom

Thanks,
Hui

Re: Tracepoints (and python)

[Tom Tromey]

>>>>> "Hui" == Hui Zhu <teawater@gmail.com> writes:

Hui> I have 3 issues with it:
Hui> 1. Not sure about upstream, but when I use it, it is hard to setup
Hui> tracepoint action with python.  How I handle it is write tracepoint
Hui> and action to a GDB source file and call soure command in python
Hui> script.

Yeah, I think you would be better off writing a new Python API to make
tracepoints directly.

Tom

Re: Tracepoints (and python)

[Doug Evans]

On Thu, Dec 20, 2012 at 6:32 AM, Tom Tromey <tromey@redhat.com> wrote:
>>>>>> "Marc" == Marc Brünink <marc@nus.edu.sg> writes:
>
> Marc> 1. Is there any technical reason why tracing is only possible on
> Marc> remote targets? Or is is "just" not implemented for anything else.
>
> I think the only reason is that nobody has done it.

This is part of gdb/gdbserver feature parity desires right?

Re: Tracepoints (and python)

[Tom Tromey]

Marc> 1. Is there any technical reason why tracing is only possible on
Marc> remote targets? Or is is "just" not implemented for anything else.

Tom> I think the only reason is that nobody has done it.

Doug> This is part of gdb/gdbserver feature parity desires right?

It is on the list on the wiki, but I think just to mention it as a
difference.  We (meaning Red Hat) plan to tackle most of the items on
that list, but I don't think this is one we are going to do.

Tom

Re: Tracepoints (and python)

[Doug Evans]

On Wed, Jan 2, 2013 at 10:28 AM, Tom Tromey <tromey@redhat.com> wrote:
> Marc> 1. Is there any technical reason why tracing is only possible on
> Marc> remote targets? Or is is "just" not implemented for anything else.
>
> Tom> I think the only reason is that nobody has done it.
>
> Doug> This is part of gdb/gdbserver feature parity desires right?
>
> It is on the list on the wiki, but I think just to mention it as a
> difference.  We (meaning Red Hat) plan to tackle most of the items on
> that list, but I don't think this is one we are going to do.

[Sorry for the resend, previous one got sent as text/html.]

Thanks.

For clarity's sake for  Marc,
what's needed isn't so much an implementation as a (pretty
significant) refactor.

Re: Tracepoints (and python)

[Marc Brünink]

On Jan 3, 2013, at 1:44 AM, Doug Evans wrote:

> On Wed, Jan 2, 2013 at 10:28 AM, Tom Tromey <tromey@redhat.com> wrote:
>> Marc> 1. Is there any technical reason why tracing is only possible on
>> Marc> remote targets? Or is is "just" not implemented for anything else.
>> 
>> Tom> I think the only reason is that nobody has done it.
>> 
>> Doug> This is part of gdb/gdbserver feature parity desires right?
>> 
>> It is on the list on the wiki, but I think just to mention it as a
>> difference.  We (meaning Red Hat) plan to tackle most of the items on
>> that list, but I don't think this is one we are going to do.
> 
> [Sorry for the resend, previous one got sent as text/html.]
> 
> Thanks.
> 
> For clarity's sake for  Marc,
> what's needed isn't so much an implementation as a (pretty
> significant) refactor.

Interesting. Actually I am trying to implement a rudimentary application tracer using the python interface. So far I have simply used breakpoints. It works like a charm but is obviously very slow. Thus, I want to switch to tracepoints. Having access to tracepoints during native debugging would be nice but is not necessary. This is a point for my personal wish list.
Another issue is the trace buffer. It would be nice to get a notification as soon as the trace buffer is full. Currently gdbserver does not sent one (assuming gdb is connected), does it?

Marc

Re: Tracepoints (and python)

[Yao Qi]

On 01/11/2013 03:29 PM, Marc Brünink wrote:
> Another issue is the trace buffer. It would be nice to get a notification as soon as the trace buffer is full. Currently gdbserver does not sent one (assuming gdb is connected), does it?

We have patches being reviewed/discussed about "the remote async 
notification on trace stopped due to trace buffer is full". 
http://sourceware.org/ml/gdb-patches/2012-12/msg00687.html

Once it goes in, users attach function to 'trace_stop' observer, and get 
notified when the trace is stopped.

-- 
Yao (齐尧)