Re: perf.data file format specification draft

[Arnaldo Carvalho de Melo]

Em Mon, May 18, 2015 at 01:21:33PM +0900, Namhyung Kim escreveu:
> Hi Arnaldo and Andi,
> 
> On Thu, May 14, 2015 at 10:11:29AM -0300, Arnaldo Carvalho de Melo wrote:
> > Em Thu, May 14, 2015 at 02:25:17PM +0200, Andi Kleen escreveu:
> > > Hi,
> > > 
> > > Since there are more and more consumers I started a description of the
> > > on-disk perf.data format. This does not replace the kernel perf event
> > > description or the manpage, but describes the parts that perf record
> > > adds.
> > > 
> > > So far it is still has some gaps and needs review. Eventually this should
> > > become part of the perf documentation.
> > > 
> > > Steven, would be good if you could fill in some details on how trace
> > > data works. 
> > 
> > I guess that would be Frédéric, and also I think this is a good
> > opportunity to remove some stuff that seem to be collected but unused,
> > namely a kallsyms copy and maybe something else.
> 
> Let me try to describe.  The tracing_data_get() does the work and it
> records the following for HEADER_TRACING_DATA in order:

Thanks for this! I'll try to write patches for the low hanging fruits,
and try to work a bit on other issues, like only collecting the format
files for events that the user asked for, perhaps even just for those
that had samples :-)

- Arnaldo
 
>  * tracing header data
>    - file magic bytes (including someone's birthday :) )
>    - file version (it's 0.5 - the only thing I can see in diff with trace-cmd
>      (version 6) is "saved-cmdline" file data which is unnecessary for perf
>    - byte order, size of long and page size of the system
> 
>  * tracing header files
>    - $tracefs/events/header_page and $tracefs/events/header_event
>    - describe ftrace raw buffer format which is unnecessary for perf unless
>      it reads the raw buffer directly (like my ftrace integration work?)
> 
>  * ftrace event files
>    - format file for each event in $tracefs/events/ftrace directory
>    - this is same as below but precedes other events, not sure why it's needed
> 
>  * (normal) event files
>    - format file for each tracepoint event
> 
>  * /proc/kallsyms
>    - for kernel symbol resolution, unnecessary for perf
> 
>  * tracing printk formats
>    - for trace_printk?  unnecessary for perf
> 
> The last two can go away at least.
> 
> Thanks,
> Namhyung
> 
> 
> > 
> > > Adrian, would be good if you could fill in the missing bits for
> > > auxtrace/itrace.
> > > Everyone else, please review and add missing information.
> > 
> > Thanks for doing this work!
> > 
> > IIRC there is a presentation written by Jiri where parts of this is
> > documented, lemme try to find it...
> > 
> > - Arnaldo
> > --
> > To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> > the body of a message to majordomo@xxxxxxxxxxxxxxx
> > More majordomo info at  http://vger.kernel.org/majordomo-info.html
> > Please read the FAQ at  http://www.tux.org/lkml/
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/