Re: bogus values of variables in userspace probes

[Michael Petlan <mpetlan@redhat.com>]

On Tue, 2015-11-24 at 12:08 -0300, Arnaldo Carvalho de Melo wrote:
> Em Tue, Nov 24, 2015 at 12:18:22PM +0100, Michael Petlan escreveu:
> > I have been playing with perf-probe tool and I found out that some bogus
> > values of a function argument are obtained by perf-record.
> > 
> > How to reproduce:
> > 
> > gcc -O0 -g -o dummy dummy.c
> > perf probe -x ./dummy --add 'isprime a'
> > perf record -e probe_dummy:isprime ./dummy
> > perf script
> > 
> > The actual output looks like the following:
> > 
> > dummy 32476 [000] 3534401.838454: probe_dummy:isprime: (400530) a=32767
> > dummy 32476 [000] 3534401.838504: probe_dummy:isprime: (400530) a=32714
> > dummy 32476 [000] 3534401.838513: probe_dummy:isprime: (400530) a=3
> > dummy 32476 [000] 3534401.838519: probe_dummy:isprime: (400530) a=4
> > dummy 32476 [000] 3534401.838525: probe_dummy:isprime: (400530) a=5
> > dummy 32476 [000] 3534401.838531: probe_dummy:isprime: (400530) a=6
> > dummy 32476 [000] 3534401.838537: probe_dummy:isprime: (400530) a=7
> > dummy 32476 [000] 3534401.838543: probe_dummy:isprime: (400530) a=13
> > dummy 32476 [000] 3534401.838561: probe_dummy:isprime: (400530) a=17
> > 
> > But if you look into the source, you can see that the function isprime()
> > is called with the following arguments:
> > 
> > int numbers[] = { 2, 3, 4, 5, 6, 7, 13, 17, 19 };
> > 
> > So the first and last ones are omitted, there are some bogus numbers instead
> > of them and all that is shifted somehow.
> > 
> > Note that when I probe for %ax register it looks correct.
> > 
> > The version of kernel/perf is 4.3.0. The architecture is x86_64.
> > Am I missing something or is it a bug?
> 
> I'm now trying to bisect that register error I found while trying to
> reproduce the problem reported above, but please consider writing a
> 'perf test' entry that does exactly what you did, i.e. have a simple
> program that you build with a known set of compile flags, then run it
> using perf probe to catch a specific function, its argument, and then
> check that the values it produces are the ones expected.
> 
> We would catch bugs in 'perf probe', the DWARF code, etc, much more
> quickly with something like that in place. :-)

Hi Arnaldo,

thanks for investigating the issue. I do not use fedora kernel, I use
what I build from git, but of course I am some versions back. So I use
4.3.0 and 4.3.0-rc5 and both do the same thing.

I have met this when writing new tests for perf-probe into the testsuite
I had been speaking about some time ago [1]. But if needed, I may add it
as a perf-test entry as you wish.

Michael

> 
> - Arnaldo