Re: [PATCH] perf tools: Fix JIT profiling on heap

[Namhyung Kim <namhyung@kernel.org>]

Hi Arnaldo,

On Fri, 17 Jan 2014 11:34:04 -0300, Arnaldo Carvalho de Melo wrote:
> Em Fri, Jan 17, 2014 at 04:44:04PM +0900, Namhyung Kim escreveu:
>> On Thu, 16 Jan 2014 20:23:27 +0000, Gaurav Jain wrote:
>> > On 1/16/14, 9:37 AM, "Arnaldo Carvalho de Melo" <acme@ghostprotocols.net>
>> > wrote:
>> >
>> >>Em Thu, Jan 16, 2014 at 10:49:31AM +0900, Namhyung Kim escreveu:
>> >>> Gaurav reported that perf cannot profile JIT program if it executes
>> >>> the code on heap.  This was because current map__new() only handle JIT
>> >>> on anon mappings - extends it to handle no_dso (heap, stack) case too.
>> >>> 
>> >>> This patch assumes JIT profiling only provides dynamic function
>> >>> symbols so check the mapping type to distinguish the case.  It'd
>> >>> provide no symbols for data mapping - if we need to support symbols on
>> >>> data mappings later it should be changed.
>> >>
>> >>But we do support symbols in data mappings, that is why we have
>> >>MAP__VARIABLE, etc, can you elaborate?
>
>> > Does perf support data mappings from perf map files? Could you please
>> > share an example of how I may be able to use this.
>
>> IIUC there's no difference between function and data mapping.  So you
>> can use same perf map file for both - in fact there's no way to use
>> different map file in a single task.  I guess perf will use it to find
>
> Do the /tmp/perf mapping has any per entry indication on the type of
> symbol it is (data, text) like ELF and kallsyms symtabs have?

Quoting Documentation/jit-interface.txt:

  Each line has the following format, fields separated with spaces:

  START SIZE symbolname

  START and SIZE are hex numbers without 0x.
  symbolname is the rest of the line, so it could contain special
  characters.

>
> It is possible for a function and a variable to have the same virtual
> address in some architectures (SPARC, iirc), that is why we have
> different MAP_ types (FUNCTION, VARIABLE) (which should really be
> renamed to TEXT, DATA).

Hmm.. didn't know that, interesting..

>
> So a 'struct map' for a data mmap should possibly point to a different
> 'dso' of the JIT /tmp/perf-... style if those maps don't have per entry
> indication of text/data.

Yes, but there's no way to do it currently.

>
>> only function symbols in function mappings and variables in data
>> mapping based on the address it accesses.
>
> Well, the lookup should figure out if the IP refers to TEXT or DATA and
> use MAP__{FUNCTION, VARIABLE} accordingly when asking for symbol
> resolution.

Right.  But in this case we cannot determine whether a symbol in the
/tmp/perf-... file is a function or variable.

>
>> What I wasn't sure is whether JIT program also produces some dynamic data.
>> And I think only perf mem command cares about data mappings, no?
>
> Well, I think it would be great to do that kind of data resolution for
> JITs the same way it is interesting to do for ELF ones :-)
>
> I need to stare harder at that patch, but with the above in mind, do we
> really have to check if the map is MAP__FUNCTION as IIRC this patch
> does?

Not sure.  For a JIT case, I guess the mapping is always executable and
we don't support data mapping yet, so it seems okay for now.

Thanks,
Namhyung