Re: [PATCH 0/6] perf/ftrace: Introduce hexadecimal type casting

[Masami Hiramatsu <mhiramat@kernel.org>]

On Thu, 18 Aug 2016 13:13:21 -0300
Arnaldo Carvalho de Melo <acme@kernel.org> wrote:

> Em Fri, Aug 19, 2016 at 01:01:43AM +0900, Masami Hiramatsu escreveu:
> > On Thu, 18 Aug 2016 11:14:42 -0300
> > Arnaldo Carvalho de Melo <acme@kernel.org> wrote:
> > 
> > > Em Thu, Aug 18, 2016 at 05:57:32PM +0900, Masami Hiramatsu escreveu:
> > > > Hi Arnaldo and Steven,
> > > > 
> > > > Here is an RFC series of hexadecimal type casting and
> > > > changing default type casting of perf and ftrace.
> > > > 
> > > > I've introduced x8,x16,x32,x64 according to previous
> > > > discussion on LKML.
> > > >   https://lkml.org/lkml/2016/8/10/339
> > > > 
> > > > This series includes not only adding hexadecimal types
> > > > (x8,x16,x32,x64), but also checking it is supported by
> > > > running kernel and keeping the backward compativility.
> > > > 
> > > > [1/6] Add hexadecimal type casting, but does not touch
> > > >    existing types like 'u8'.
> > > > [2/6] Show the supported types on README of ftrace so
> > > >    that user application (e.g. perf) can check that.
> > > > [3/6] Add a type availability check to perf-probe.
> > > > [4/6] Add hexadecimal prefix support to perf-probe if
> > > >    it is supported by the kernel. 
> > > > [5/6] Change the perf-probe default type casting for
> > > >    unsigned type to hexadecimal (for backward compatibility)
> > > > [6/6] Change ftrace's 'uNN' to show value in decimal
> > > >    and use 'xNN' by default (for backward compatibility)
> > > > 
> > > > This way, we can also add "octal" type, pointer type,
> > > > and "character" type etc. and perf can check whether
> > > > the kernel supports it or not. :)
> > > 
> > > But this requires a kernel update... If we do it all in the tooling
> > > side, no kernel changes are required _and_ newer tools will work with
> > > older kernels, as this is just a formatting issue, the value is there
> > > and from its format one can infer its value, it is not even necessary to
> > > look at its "type".
> > > 
> > > I understand this is necessary for ftrace, because the pretty printer is
> > > in the kernel, but I don't see why we would prevent tooling from doing
> > > this pretty printing work and make it support any kernel.
> > > 
> > > I.e. no need at all for checking if the kernel supports anything, just
> > > pretty print it.
> > 
> > That's should be handled by libtraceevent. And if you need just
> > a pretty printing, you can do it in python/perl script via perf-script.
> > Anyway, to see the event parameters via perf tools, we have to use
> > perf-script with/without scripts. As you know, perf-script without
> > scripts, the output format depends on libtraceevent. And it seems that
> > the libtraceevent output depends on ftrace event "format" file.
>  
> > And also as you can see in Naohiro's report　( https://lkml.org/lkml/2016/8/5/191 )
> > he was using perf-probe with ftrace trace_pipe interface, which was my
> > expected usecase. In that case we have to change ftrace side to
> > support various pretty printing.
> 
> Sure, if one wants to have pretty printing via trace_pipe, then a kernel
> reboot is needed, I'm talking about those who don't want to or can't
> reboot their machines :-)
> 
> But anyway, your patches allow seeing those events in decimal for people
> that can reboot their machine when that wasn't possible without extra
> scripting via perl/python, a clear advance!
> 
> It remains to be coded a way to achieve the same result without
> requiring a kernel reboot, i.e. a tooling only change.

Hmm, one solution for user space tool is to identify the probed event
from event name, and decide output format based on the format locally
stored, like cached entries. Actually, current perf-probe to perf-script
execution chain lacks such kind of information, that is a problem.

 DWARF -> perf-probe -> ftrace -> perf-record -> perf-script

perf-probe can read DWARF(debuginfo) so it can handle a variety of types,
but between perf-probe to ftrace, it lacks type information and all
types falls into s8/16/32/64, u8/16/32/64, string and bitfield.
So after that process, perf-record treats data as a raw binary,
and perf-script tries to parse that binary based on ftrace's event
format. 

But if we can pass the data format in perf-cache entries like below

 DWARF -> perf-probe -> ftrace -> perf-record -+-> perf-script
                 +-------> probe-cache entry --+

then perf-probe can pass correct format to perf-script.

But anyway, I will update ftrace to add other types since that will
help ftrace users.
 
Thanks,

-- 
Masami Hiramatsu <mhiramat@kernel.org>