[BUG] perf top: -z option does not work

[BUG] perf top: -z option does not work

[Stephane Eranian]

Hi,


My understanding is that the -z option is used to only print a profile
since the last refresh. So if I have a refresh of 5s, then it prints the profile
based on the samples accumulated over the last 5 seconds.

The Z mode used to be available interactively. Nowadays, it seems only
avail from the cmdline. But it does not work.

I run a simple test.
$ sudo perf top -z

Then for 5s I run a cycle-burning noploop program.
It shows up at the top. But when it terminates, the program  still stays
at the top of the profile for a long time. This is not what I'd expect.
The noploop line should disappear in the next couple of refreshes after
the program has terminated.

I am using tip.git and seeing the problem. But it seems, it's been there for a
long time.

Any idea what's wrong?

Thanks.

Re: [BUG] perf top: -z option does not work

[Namhyung Kim]

Hi Stephane,

On Tue, 12 Aug 2014 06:44:34 +0200, Stephane Eranian wrote:
> Hi,
>
>
> My understanding is that the -z option is used to only print a profile
> since the last refresh. So if I have a refresh of 5s, then it prints the profile
> based on the samples accumulated over the last 5 seconds.

Yep, that's what I understand about the -z behavior too.

>
> The Z mode used to be available interactively. Nowadays, it seems only
> avail from the cmdline. But it does not work.

Hmm.. it seems the stdio supports 'z' key but TUI don't.

>
> I run a simple test.
> $ sudo perf top -z
>
> Then for 5s I run a cycle-burning noploop program.
> It shows up at the top. But when it terminates, the program  still stays
> at the top of the profile for a long time. This is not what I'd expect.
> The noploop line should disappear in the next couple of refreshes after
> the program has terminated.
>
> I am using tip.git and seeing the problem. But it seems, it's been there for a
> long time.
>
> Any idea what's wrong?

Looking at the code, it only zero out the annotation info but hist
entries.  I guess we need to check the flag and throw out existing
entries instead of decaying.

Also I wonder about the order of decaying - shouldn't it be decayed
before processing current entries?  It seems current code processes
current entries first and then decays...

I'll prepare patches for this soon.

Thanks,
Namhyung

Re: [BUG] perf top: -z option does not work

[Arnaldo Carvalho de Melo]

Em Tue, Aug 12, 2014 at 05:09:17PM +0900, Namhyung Kim escreveu:
> Hi Stephane,
> 
> On Tue, 12 Aug 2014 06:44:34 +0200, Stephane Eranian wrote:
> > Hi,
> >
> >
> > My understanding is that the -z option is used to only print a profile
> > since the last refresh. So if I have a refresh of 5s, then it prints the profile
> > based on the samples accumulated over the last 5 seconds.
> 
> Yep, that's what I understand about the -z behavior too.
> 
> >
> > The Z mode used to be available interactively. Nowadays, it seems only
> > avail from the cmdline. But it does not work.
> 
> Hmm.. it seems the stdio supports 'z' key but TUI don't.

Problem introduced long ago when I moved 'top' to use the same hists
browser as 'report', to have all the zooming, --sort, etc goodies.

I see that Namhyung proposed a patch to fix that, will review it,
thanks!

- Arnaldo
 
> >
> > I run a simple test.
> > $ sudo perf top -z
> >
> > Then for 5s I run a cycle-burning noploop program.
> > It shows up at the top. But when it terminates, the program  still stays
> > at the top of the profile for a long time. This is not what I'd expect.
> > The noploop line should disappear in the next couple of refreshes after
> > the program has terminated.
> >
> > I am using tip.git and seeing the problem. But it seems, it's been there for a
> > long time.
> >
> > Any idea what's wrong?
> 
> Looking at the code, it only zero out the annotation info but hist
> entries.  I guess we need to check the flag and throw out existing
> entries instead of decaying.
> 
> Also I wonder about the order of decaying - shouldn't it be decayed
> before processing current entries?  It seems current code processes
> current entries first and then decays...
> 
> I'll prepare patches for this soon.
> 
> Thanks,
> Namhyung