Messages in this thread | ![/](/images/icornerl.gif) | | Date | Thu, 31 Dec 2009 08:29:38 -0200 | From | Arnaldo Carvalho de Melo <> | Subject | Re: [PATCH 3/3] perf tools: adjust symbol address |
| |
Em Thu, Dec 31, 2009 at 10:59:35AM +0800, Xiao Guangrong escreveu: > > > Arnaldo Carvalho de Melo wrote: > > Em Wed, Dec 30, 2009 at 11:18:55AM +0800, Xiao Guangrong escreveu: > >> Using relocation offset adjust symbol address if we get > >> kernel symbol name form elf file > >> > >> diff --git a/tools/perf/util/symbol.c b/tools/perf/util/symbol.c > >> index 79ca6a0..5b58d34 100644 > >> --- a/tools/perf/util/symbol.c > >> +++ b/tools/perf/util/symbol.c > >> @@ -19,6 +19,18 @@ > >> #define NT_GNU_BUILD_ID 3 > >> #endif > >> > >> +static s32 relocate_offset; > >> +void update_relocate_offset(s32 offset) > >> +{ > >> + relocate_offset = offset; > >> +} > >> + > >> +static inline void update_kernel_address(GElf_Sym *sym, bool kernel) > >> +{ > >> + if (kernel) > >> + sym->st_value += relocate_offset; > >> +}
> > This should be done on the dso level, we may process multiple kernels, > > some with relocation, some without.
> Sorry, what does the 'multiple kernels' means?
perf diff? I think you are combining the 'perf record' and 'perf report' steps as a single action...
> vmlinux* not know it's relocation or not, and we just handle the _specific_ > kernel at the time which match the current system.
'perf record' in one machine + 'perf report' in another should be possible.
> And i think adjust symbol address on the dso level is not a good idea since > we hardly know it's a module address or kernel address on dso level, i.e. > hardly to know whether the address need adjust.
yeah, as I contininued below, its not the right level, for other reasons.
> > Humm, even at the _map_ level, because then we can just use the normal > > map_ip mechanism, this time the not using the identity map stuff, i.e. > > we have to get an IP from some event, then subtract the relocate_offset > > that will be in map->start. > > > > The 'inject' event is the first event in perf.data since we inject it at the > first counter file and inject 'relocation offset' before enable event, i.e. > before getting symbol address we have got the 'relocation offset' already.
There is no need for kernel changes, as the discussion showed.
- Arnaldo
| ![\](/images/icornerr.gif) |