On Fri, Aug 07, 2009 at 08:38:56AM -0700, Manfred M?cke wrote:
> Hello,
>
> I have a fortran program and want to monitor entry/return events of a
> specific function. No big thing (I thought). The function is called DCNTRL in
> the source code. Checking the binary gives:
>
> > nm myprog | grep FUNC | grep -i DCNTRL
> [15968] | 6564704| 30193|FUNC |GLOB |0 |12 |dcntrl_
>
> OK, its internal name is dcntrl_. Let's instrument the respective pid
> entry/return probes:
>
> > dtrace -n 'pid$target::dcntrl_:entry' -n 'pid$target::dcntrl_:return' -c
> myprog
> dtrace: invalid probe specifier pid$target::dcntrl_:return: probe
> description pid17186::dcntrl_:return does not match any probes
>
> Huh, DTrace does not find the return probe!? But I'm sure the function exist,
> so let's put some pressure on DTrace (-Z):
>
> > dtrace -Z -n 'pid$target::dcntrl_:entry' -n 'pid$target::dcntrl_:return'
> -c myprog
> dtrace: description 'pid$target::dcntrl*:entry' matched 1 probe
> dtrace: description 'pid$target::dcntrl*:return' matched 0 probes
>
> First question: How comes DTrace can instrument the entry probe, but
> not the return probe of that function?
If dtrace thinks something fishy is going on, it can refuse to instrument
the return sites. This is mainly due to issues with jump-tables.
If you set DTRACE_DEBUG= in your environment, is there anything of interest?
Look for "grabbed pid NNNN" as a starting place.
> Now let's count events a the instrumented probes:
>
> > dtrace -Z -n 'pid$target::dcntrl_:entry
> {...@f[probefunc,probename]=count()}' -n 'pid$target::dcntrl_:return
> {...@f[probefunc,probename]=count()}' -c myprog
>
> dcntrl_ entry
> 1
>
> Hm, seems like that function was only called once. Fact is, it was
> called a couple of times (verified), but the probe seems only to fire
> the first time.
>
> I conclude that I'm unable to both instrument and observe the
> function. Does anyone have an idea, how I could proceed?
>
> Thanks for your help, Manfred
>
> PS: I explicitly switched inlining off for that function.
> (-inline=no%dcntrl). Just for reference, I attach the first and last
> lines of the function:
>
> > mdb myprog
> > dcntrl_::dis
> dcntrl_: pushq %rbp
> dcntrl_+1: movq %rsp,%rbp
> dcntrl_+4: pushq %rbx
> dcntrl_+5: pushq %r12
> dcntrl_+7: pushq %r13
> dcntrl_+9: pushq %r14
> dcntrl_+0xb: pushq %r15
> [..]
> dcntrl_+0x75e6: popq %r15
> dcntrl_+0x75e8: popq %r14
> dcntrl_+0x75ea: popq %r13
> dcntrl_+0x75ec: popq %r12
> dcntrl_+0x75ee: popq %rbx
> dcntrl_+0x75ef: leave
> dcntrl_+0x75f0: ret
One workaround might be to instrument pidNNNN::dcntrl:75f0
but that will change with recompilation. (And you'd need to get any other rets)
Cheers,
- jonathan
_______________________________________________
dtrace-discuss mailing list
[email protected]
