Debugging agent library

Debugging agent library

[Suchakra Sharma]

Hi,

We tried to find out how to use the debugging agent library here [1] but
were unsuccessful. As of now, we have built a test app linked to dagent
which creates a socket file and expects commands. We wish to use it with
GDB as well as other applications to primarily objective of tracing
using GDB's fast tracepoint infra. Any help in this regard will be
appreciated. 

[1] https://github.com/MentorEmbedded/Debug-Agent-Library/wiki

Regards,
Suchakra & Simon

--
Suchakrapani Datt Sharma
Laboratorie DORSAL
École Polytechnique de Montréal

Re: Debugging agent library

[Simon Marchi]

Hi Yao,

On 7 May 2013 03:33, Yao Qi <yao@codesourcery.com> wrote:
> Hello,
> Suchakra,
>
>
> On 05/07/2013 03:21 AM, Suchakra Sharma wrote:
>>
>> Hi,
>>
>> We tried to find out how to use the debugging agent library here [1] but
>> were unsuccessful. As of now, we have built a test app linked to dagent
>
>
> It would be nice if you post some details about how you use dagent and the
> error message you got.
>
>
>> which creates a socket file and expects commands. We wish to use it with
>> GDB as well as other applications to primarily objective of tracing
>> using GDB's fast tracepoint infra. Any help in this regard will be
>> appreciated.
>>
>> [1] https://github.com/MentorEmbedded/Debug-Agent-Library/wiki
>>
>
> I checked out the wiki and realize that the existing documentation has an
> assumption that readers are much familiar with GDB and GDB tracepoint.  The
> session below shows how to install fast tracepoint through dagent.
>
> In one console, start the program in GDBserver,
>
> $ LD_PRELOAD=/install/lib/libdagent.so ./gdbserver/gdbserver :1234
> ./testsuite/gdb.base/break
>
> In the other console, start GDB and connect to GDBserver,
>
> $ ./gdb ./testsuite/gdb.base/break
> (gdb) target remote :1234
> (gdb) b main
> (gdb) c
>
> (gdb) b marker4  // set breakpoint to make sure program doesn't exit
> (gdb) ftrace *0x0804851e // set fast tracepoint at certain address
> (gdb) set agent on   // enable agent
> (gdb) tstart         // start tracing

tstart here gives me the error:

(gdb) tstart
Target returns error code '.In-process agent library not loaded in
process.  Fast and static tracepoints unavailable.'

I compiled my small app using (libdagent.so is already in my ld search path):
$ gcc test.c -Wall -g -ldagent -O0

I run the gdbserver using:
$ LD_PRELOAD=/usr/local/lib/libdagent.so gdbserver :1234 ./a.out

Then I use the same commands as you have shown in your example. info
sharedlibrary shows that the agent is loaded:

(gdb) info sharedlibrary
From                To                  Syms Read   Shared Object Library
0x00007ffff7dddb00  0x00007ffff7df685b  Yes (*)     /lib64/ld-linux-x86-64.so.2
0x00007ffff7bd38c0  0x00007ffff7bd8e38  Yes         /usr/local/lib/libdagent.so
0x00007ffff785ba20  0x00007ffff797c72c  Yes (*)     /lib64/libc.so.6
0x00007ffff7625660  0x00007ffff7630eb8  Yes (*)     /lib64/libpthread.so.0

Is there something I am missing here?

> (gdb) c
> (gdb) tstop          // stop tracing
> (gdb) tfind          // examine the trace frames tracepoint collects
> Found trace frame 0, tracepoint 3
> #0  main (argc=<unavailable>, argv=<unavailable>, envp=<unavailable>) at
> ../../../../git/gdb/testsuite/gdb.base/break.c:94
> 94          marker1 ();  /* set breakpoint 11 here */
>
> Let me know if you get something else or obviously wrong.
>
> --
> Yao (齐尧)

Also, is it possible to write our own application to control tracing,
instead of using GDB? I guess we would need to talk to the agent using
the IPA protocol, is that right?

Thanks !

Simon

Re: Debugging agent library

[Yao Qi]

On 05/08/2013 02:49 AM, Simon Marchi wrote:
> tstart here gives me the error:
>
> (gdb) tstart
> Target returns error code '.In-process agent library not loaded in
> process.  Fast and static tracepoints unavailable.'
>
> I compiled my small app using (libdagent.so is already in my ld search path):
> $ gcc test.c -Wall -g -ldagent -O0
>
> I run the gdbserver using:
> $ LD_PRELOAD=/usr/local/lib/libdagent.so gdbserver :1234 ./a.out

If libdagent is linked to your program explicitly, LD_PRELOAD is not 
needed any more.  You have to either compile your program without 
-ldagent and use LD_PRELOAD on run or compile your program with -ldagent 
and run your program normally.

>
> Then I use the same commands as you have shown in your example. info
> sharedlibrary shows that the agent is loaded:
>
> (gdb) info sharedlibrary
>  From                To                  Syms Read   Shared Object Library
> 0x00007ffff7dddb00  0x00007ffff7df685b  Yes (*)     /lib64/ld-linux-x86-64.so.2
> 0x00007ffff7bd38c0  0x00007ffff7bd8e38  Yes         /usr/local/lib/libdagent.so
> 0x00007ffff785ba20  0x00007ffff797c72c  Yes (*)     /lib64/libc.so.6
> 0x00007ffff7625660  0x00007ffff7630eb8  Yes (*)     /lib64/libpthread.so.0
>
> Is there something I am missing here?
>

If the suggestions above don't work for you, you probably have to turn 
on some debugging outputs.  You can start gdbserver with option 
"--debug", and further, hack the variable "debug_agent" to 1 in 
gdb/common/agent.c.

> Also, is it possible to write our own application to control tracing,
> instead of using GDB? I guess we would need to talk to the agent using
> the IPA protocol, is that right?

Tracing control involves two sides, GDB and the agent.  IPA protocol is 
used to control the agent, however, in order to finish/control tracing, 
GDB side should be controlled as well.  Your tool can get the trace data 
from agent, but these data needs GDB interpretation, map the address to 
a symbol, for example.  I am afraid you can't use your own application 
to do tracing with agent.

On the other hand, I am curious on the reasons to write your own tool to 
control tracing instead of using GDB.  Tracing is interesting to GDB, 
and you can see how much progress GDB made on tracing in the past 
several years.  GDB is extensible on both its c code and python script, 
and you can extend it for your purpose.  Last by not least, 
contributions are always welcome, and people here are friendly and nice 
to answer questions, AFAICS.

-- 
Yao (齐尧)

Re: Debugging agent library

[Suchakra Sharma]

Hi Yao,

> If the suggestions above don't work for you, you probably have to turn 
> on some debugging outputs.  You can start gdbserver with option 
> "--debug", and further, hack the variable "debug_agent" to 1 in 
> gdb/common/agent.c.

Thanks for your detailed reply. We were able to rectify the issue. We
tried with a clean install of gdb and libdagent on another machine which
did not show any issue. The problem was probably because we tinkered
around the dagent code a bit before using it. All works well now.

> > Also, is it possible to write our own application to control tracing,
> > instead of using GDB? I guess we would need to talk to the agent using
> > the IPA protocol, is that right?
> 
> Tracing control involves two sides, GDB and the agent.  IPA protocol is 
> used to control the agent, however, in order to finish/control tracing, 
> GDB side should be controlled as well.  Your tool can get the trace data 
> from agent, but these data needs GDB interpretation, map the address to 
> a symbol, for example.  I am afraid you can't use your own application 
> to do tracing with agent.
> 
> On the other hand, I am curious on the reasons to write your own tool to 
> control tracing instead of using GDB.  Tracing is interesting to GDB, 
> and you can see how much progress GDB made on tracing in the past 
> several years.  GDB is extensible on both its c code and python script, 
> and you can extend it for your purpose.  Last by not least, 
> contributions are always welcome, and people here are friendly and nice 
> to answer questions, AFAICS.

Indeed, dynamic tracing with GDB is getting more awesome day by day
(this view is reinforced by some tests I ran in-house which really show
how powerful and fast it is) A you know LTTng/UST already has quite fast
static tracing features, so our main aim is to get dynamic tracing
infrastructure of GDB in LTTng/UST for fast dynamic user space tracing
in an elegant way. We are not 100% clear how it shall be achieved at
this moment but we are exploring all avenues for now. At first glance it
looked as if we could  do it using this library coupled with GDB
(controlled through LTTng/UST acting as our tool). Your ideas are
welcome and we surely shall contribute once we get a good knack of how
to extend GDB's functionality for our use.

Regards,
Suchakra & Simom

--
Suchakrapani Datt Sharma
Laboratorie DORSAL
École Polytechnique de Montréal

Re: Debugging agent library

[Yao Qi]

On 05/09/2013 03:41 AM, Suchakra Sharma wrote:
> Thanks for your detailed reply. We were able to rectify the issue. We
> tried with a clean install of gdb and libdagent on another machine which
> did not show any issue. The problem was probably because we tinkered
> around the dagent code a bit before using it. All works well now.
>

Good.

> Indeed, dynamic tracing with GDB is getting more awesome day by day
> (this view is reinforced by some tests I ran in-house which really show
> how powerful and fast it is) A you know LTTng/UST already has quite fast
> static tracing features, so our main aim is to get dynamic tracing
> infrastructure of GDB in LTTng/UST for fast dynamic user space tracing
> in an elegant way. We are not 100% clear how it shall be achieved at
> this moment but we are exploring all avenues for now. At first glance it
> looked as if we could  do it using this library coupled with GDB
> (controlled through LTTng/UST acting as our tool). Your ideas are
> welcome and we surely shall contribute once we get a good knack of how
> to extend GDB's functionality for our use.

GDB/GDBserver is able to control UST static tracepoint through agent, 
but UST is 0.11 and URCU is 0.5.3, which is a little bit old.  In UST 
0.12 release, the GDB support is removed unfortunately.  Probably you 
can have a look at testsuite/gdb.trace/strace.exp to see how 
GDB/GDBserver interact with UST static tracepoints.

-- 
Yao (齐尧)