Does perf inject preserve call chain data?

Does perf inject preserve call chain data?

[Tim Chen]

Arnaldo,

Wonder if perf inject munged the call chain data from perf record?

When I do
perf record -a -g -o - sleep 5 |perf inject -v -b  | perf report  -g -i -
I get a complain from perf report

Warning:
Selected -g but no callchain data. Did you call 'perf record' without -g?

Thanks.

Tim

Re: Does perf inject preserve call chain data?

[David Ahern]

On 5/22/12 12:49 PM, Tim Chen wrote:
> Arnaldo,
>
> Wonder if perf inject munged the call chain data from perf record?
>
> When I do
> perf record -a -g -o - sleep 5 |perf inject -v -b  | perf report  -g -i -
> I get a complain from perf report
>
> Warning:
> Selected -g but no callchain data. Did you call 'perf record' without -g?
>

I believe the root cause is that the event list is not written to the 
pipe. Using Arnaldo's git repo as of today (has fixes from Stephane) 
this hack shows the info is there and can be processed.


diff --git a/tools/perf/builtin-report.c b/tools/perf/builtin-report.c
index d58e414..60f3f17 100644
--- a/tools/perf/builtin-report.c
+++ b/tools/perf/builtin-report.c
@@ -256,7 +257,7 @@ static int perf_report__setup_sample_type(struct 
perf_report *rep)
                     "'perf record' without -g?\n");
             return -EINVAL;
         }
-       if (symbol_conf.use_callchain) {
+       if (!self->fd_pipe && symbol_conf.use_callchain) {
             ui__warning("Selected -g but no callchain data. Did "
                     "you call 'perf record' without -g?\n");
             return -1;

A proper solution is needed.

David

Re: Does perf inject preserve call chain data?

[Arnaldo Carvalho de Melo]

Em Tue, May 22, 2012 at 03:58:40PM -0600, David Ahern escreveu:
> On 5/22/12 12:49 PM, Tim Chen wrote:
> >Arnaldo,
> >
> >Wonder if perf inject munged the call chain data from perf record?
> >
> >When I do
> >perf record -a -g -o - sleep 5 |perf inject -v -b  | perf report  -g -i -
> >I get a complain from perf report
> >
> >Warning:
> >Selected -g but no callchain data. Did you call 'perf record' without -g?
> >
> 
> I believe the root cause is that the event list is not written to
> the pipe. Using Arnaldo's git repo as of today (has fixes from
> Stephane) this hack shows the info is there and can be processed.

How things stand here? Tim, did this fix things for you? If so, David,
please resend with a proper subject line and a Tested-by: Tim, ok?

Thanks,

- Arnaldo
 
> 
> diff --git a/tools/perf/builtin-report.c b/tools/perf/builtin-report.c
> index d58e414..60f3f17 100644
> --- a/tools/perf/builtin-report.c
> +++ b/tools/perf/builtin-report.c
> @@ -256,7 +257,7 @@ static int perf_report__setup_sample_type(struct
> perf_report *rep)
>                     "'perf record' without -g?\n");
>             return -EINVAL;
>         }
> -       if (symbol_conf.use_callchain) {
> +       if (!self->fd_pipe && symbol_conf.use_callchain) {
>             ui__warning("Selected -g but no callchain data. Did "
>                     "you call 'perf record' without -g?\n");
>             return -1;
> 
> A proper solution is needed.
> 
> David

Re: Does perf inject preserve call chain data?

[Tim Chen]

On Fri, 2012-06-08 at 11:38 -0300, Arnaldo Carvalho de Melo wrote:
> Em Tue, May 22, 2012 at 03:58:40PM -0600, David Ahern escreveu:
> > On 5/22/12 12:49 PM, Tim Chen wrote:
> > >Arnaldo,
> > >
> > >Wonder if perf inject munged the call chain data from perf record?
> > >
> > >When I do
> > >perf record -a -g -o - sleep 5 |perf inject -v -b  | perf report  -g -i -
> > >I get a complain from perf report
> > >
> > >Warning:
> > >Selected -g but no callchain data. Did you call 'perf record' without -g?
> > >
> > 
> > I believe the root cause is that the event list is not written to
> > the pipe. Using Arnaldo's git repo as of today (has fixes from
> > Stephane) this hack shows the info is there and can be processed.
> 
> How things stand here? Tim, did this fix things for you? If so, David,
> please resend with a proper subject line and a Tested-by: Tim, ok?
> 
> Thanks,
> 
> - Arnaldo
>  

It didn't fix things for me when I was testing on a westmere-ex machine.
I got

# ./perf record -a -g -o - sleep 5 |perf inject -v -b  | perf report -g  -i -
incompatible file format
incompatible file formatincompatible file formatselected -g but no callchain data. Did you call perf record without -g?

Testing was done on 3.5.0-rc1.

Tim

> > 
> > diff --git a/tools/perf/builtin-report.c b/tools/perf/builtin-report.c
> > index d58e414..60f3f17 100644
> > --- a/tools/perf/builtin-report.c
> > +++ b/tools/perf/builtin-report.c
> > @@ -256,7 +257,7 @@ static int perf_report__setup_sample_type(struct
> > perf_report *rep)
> >                     "'perf record' without -g?\n");
> >             return -EINVAL;
> >         }
> > -       if (symbol_conf.use_callchain) {
> > +       if (!self->fd_pipe && symbol_conf.use_callchain) {
> >             ui__warning("Selected -g but no callchain data. Did "
> >                     "you call 'perf record' without -g?\n");
> >             return -1;
> > 
> > A proper solution is needed.
> > 
> > David

Re: Does perf inject preserve call chain data?

[David Ahern]

On 6/8/12 10:44 AM, Tim Chen wrote:
> On Fri, 2012-06-08 at 11:38 -0300, Arnaldo Carvalho de Melo wrote:
>> Em Tue, May 22, 2012 at 03:58:40PM -0600, David Ahern escreveu:
>>> On 5/22/12 12:49 PM, Tim Chen wrote:
>>>> Arnaldo,
>>>>
>>>> Wonder if perf inject munged the call chain data from perf record?
>>>>
>>>> When I do
>>>> perf record -a -g -o - sleep 5 |perf inject -v -b  | perf report  -g -i -
>>>> I get a complain from perf report
>>>>
>>>> Warning:
>>>> Selected -g but no callchain data. Did you call 'perf record' without -g?
>>>>
>>>
>>> I believe the root cause is that the event list is not written to
>>> the pipe. Using Arnaldo's git repo as of today (has fixes from
>>> Stephane) this hack shows the info is there and can be processed.
>>
>> How things stand here? Tim, did this fix things for you? If so, David,
>> please resend with a proper subject line and a Tested-by: Tim, ok?
>>
>> Thanks,
>>
>> - Arnaldo
>>
>
> It didn't fix things for me when I was testing on a westmere-ex machine.
> I got
>
> # ./perf record -a -g -o - sleep 5 |perf inject -v -b  | perf report -g  -i -
> incompatible file format
> incompatible file formatincompatible file formatselected -g but no callchain data. Did you call perf record without -g?
>
> Testing was done on 3.5.0-rc1.

Is ./ needed in front of the second and third invocations are do you 
have ./ in the PATH?

David

Re: Does perf inject preserve call chain data?

[Tim Chen]

On Fri, 2012-06-08 at 11:13 -0600, David Ahern wrote:
> On 6/8/12 10:44 AM, Tim Chen wrote:
> > On Fri, 2012-06-08 at 11:38 -0300, Arnaldo Carvalho de Melo wrote:
> >> Em Tue, May 22, 2012 at 03:58:40PM -0600, David Ahern escreveu:
> >>> On 5/22/12 12:49 PM, Tim Chen wrote:
> >>>> Arnaldo,
> >>>>
> >>>> Wonder if perf inject munged the call chain data from perf record?
> >>>>
> >>>> When I do
> >>>> perf record -a -g -o - sleep 5 |perf inject -v -b  | perf report  -g -i -
> >>>> I get a complain from perf report
> >>>>
> >>>> Warning:
> >>>> Selected -g but no callchain data. Did you call 'perf record' without -g?
> >>>>
> >>>
> >>> I believe the root cause is that the event list is not written to
> >>> the pipe. Using Arnaldo's git repo as of today (has fixes from
> >>> Stephane) this hack shows the info is there and can be processed.
> >>
> >> How things stand here? Tim, did this fix things for you? If so, David,
> >> please resend with a proper subject line and a Tested-by: Tim, ok?
> >>
> >> Thanks,
> >>
> >> - Arnaldo
> >>
> >
> > It didn't fix things for me when I was testing on a westmere-ex machine.
> > I got
> >
> > # ./perf record -a -g -o - sleep 5 |perf inject -v -b  | perf report -g  -i -
> > incompatible file format
> > incompatible file formatincompatible file formatselected -g but no callchain data. Did you call perf record without -g?
> >
> > Testing was done on 3.5.0-rc1.
> 
> Is ./ needed in front of the second and third invocations are do you 
> have ./ in the PATH?
> 
> David

Thanks for catching my dumb mistake.  Things work properly now.

Tim

Re: Does perf inject preserve call chain data?

[Andi Kleen]

> > >
> > > # ./perf record -a -g -o - sleep 5 |perf inject -v -b  | perf report -g  -i -

Since this is a common pattern with files, but files are often
slow and too big, it would be nice if perf had a 
standard alias that started this pipeline.

-Andi


-- 
ak@linux.intel.com -- Speaking for myself only

Re: Does perf inject preserve call chain data?

[Arnaldo Carvalho de Melo]

Em Fri, Jun 08, 2012 at 12:52:15PM -0700, Andi Kleen escreveu:
> > > > # ./perf record -a -g -o - sleep 5 |perf inject -v -b  | perf report -g  -i -
 
> Since this is a common pattern with files, but files are often slow
> and too big, it would be nice if perf had a standard alias that
> started this pipeline.

I think this is just a variation for perf top, no? I.e. something like:

# perf top -a -G sleep 5

Where it would figure out that since a workload was specified, it should
just not be refreshing the screen every N seconds but wait till the
workload to finish and then present the results, without setting up any
perf.data files.

It could as well show intermediate results every N seconds, but not decay
the older samples, which should be even nicer.

For workloads where there is output it could spawn another term, in the
background perhaps, that the user could switch to to see the samples
collected while the workload runs.

That looks like a step in the direction of more integration of the
tools, i.e. record + report (top being report+refreshes) + annotate.

- Arnaldo

Re: Does perf inject preserve call chain data?

[Andi Kleen]

> # perf top -a -G sleep 5
>
> Where it would figure out that since a workload was specified, it should
> just not be refreshing the screen every N seconds but wait till the
> workload to finish and then present the results, without setting up any
> perf.data files.

But top does not generate email'able ASCII format like perf report does, 
doesn't it?



-Andi

Re: Does perf inject preserve call chain data?

[David Ahern]

On 6/11/12 12:43 PM, Andi Kleen wrote:
>
>
>> # perf top -a -G sleep 5
>>
>> Where it would figure out that since a workload was specified, it should
>> just not be refreshing the screen every N seconds but wait till the
>> workload to finish and then present the results, without setting up any
>> perf.data files.
>
> But top does not generate email'able ASCII format like perf report does,
> doesn't it?


perf top --stdio ...

Re: Does perf inject preserve call chain data?

[Arnaldo Carvalho de Melo]

Em Mon, Jun 11, 2012 at 11:43:30AM -0700, Andi Kleen escreveu:
> ># perf top -a -G sleep 5
> >
> >Where it would figure out that since a workload was specified, it should
> >just not be refreshing the screen every N seconds but wait till the
> >workload to finish and then present the results, without setting up any
> >perf.data files.
> 
> But top does not generate email'able ASCII format like perf report
> does, doesn't it?

http://git.kernel.org/?p=linux/kernel/git/acme/linux.git;a=commit;h=a78e5b82e99d8d284afc012ef015ef3332153aaf

perf hists browser: Implement printing snapshots to files

To avoid having to resort to --stdio, that expands everything, instead
allow the user to go on expanding the relevant callchains and then press
'P' to print that view.

As the hists browser is used for both static (report) and dynamic (top)
views, it prints to a 'perf.hists.N' sequence, i.e. multiple snapshots
can be taken in report and top.

</>

On a v3.6 kernel near you, together with:

http://git.kernel.org/?p=linux/kernel/git/acme/linux.git;a=commit;h=bd622d64e4bb0d10e8f0c2df4de9303ee415ffae

:-)

- Arnaldo

Re: Does perf inject preserve call chain data?

[Andi Kleen]

> http://git.kernel.org/?p=linux/kernel/git/acme/linux.git;a=commit;h=a78e5b82e99d8d284afc012ef015ef3332153aaf

Thanks. That would work I suppose.  I think I would still prefer a single
command over a bunch of options, but having options is not too bad.

-Andi

Re: Does perf inject preserve call chain data?

[Arnaldo Carvalho de Melo]

Em Mon, Jun 11, 2012 at 03:11:35PM -0700, Andi Kleen escreveu:
> >http://git.kernel.org/?p=linux/kernel/git/acme/linux.git;a=commit;h=a78e5b82e99d8d284afc012ef015ef3332153aaf
> 
> Thanks. That would work I suppose.  I think I would still prefer a single
> command over a bunch of options, but having options is not too bad.

Not a bunch, just one option.

And only if you feel like, at that point in time, to snapshot it to send
it over e-mail.

- Arnaldo

Re: Does perf inject preserve call chain data?

[David Ahern]

On 6/8/12 11:17 AM, Tim Chen wrote:
> On Fri, 2012-06-08 at 11:13 -0600, David Ahern wrote:
>> On 6/8/12 10:44 AM, Tim Chen wrote:
>>> On Fri, 2012-06-08 at 11:38 -0300, Arnaldo Carvalho de Melo wrote:
>>>> Em Tue, May 22, 2012 at 03:58:40PM -0600, David Ahern escreveu:
>>>>> On 5/22/12 12:49 PM, Tim Chen wrote:
>>>>>> Arnaldo,
>>>>>>
>>>>>> Wonder if perf inject munged the call chain data from perf record?
>>>>>>
>>>>>> When I do
>>>>>> perf record -a -g -o - sleep 5 |perf inject -v -b  | perf report  -g -i -
>>>>>> I get a complain from perf report
>>>>>>
>>>>>> Warning:
>>>>>> Selected -g but no callchain data. Did you call 'perf record' without -g?
>>>>>>
>>>>>
>>>>> I believe the root cause is that the event list is not written to
>>>>> the pipe. Using Arnaldo's git repo as of today (has fixes from
>>>>> Stephane) this hack shows the info is there and can be processed.
>>>>
>>>> How things stand here? Tim, did this fix things for you? If so, David,
>>>> please resend with a proper subject line and a Tested-by: Tim, ok?

Pipeline is broken by this commit in your urgent (but not yet pushed) tree:

commit 47832ad3e101e470a90fab9b4915549aaee03662
Author: Dmitry Antipov <dmitry.antipov@linaro.org>
Date:   Fri Jun 8 16:23:27 2012 +0400

     perf report: Fix event name reporting in cross analysis

     Use trace_find_event to find event name before looking through /sys
     files. This helps 'perf report' to show real event names instead of
     'unknown:unknown' when processing perf.data recorded on another 
machine.


pevent might be NULL and pevent_find_event dereferences it causing a crash.

David