Re: [PROBLEM] perf report is broken in latest linus git

[Peter Zijlstra <peterz@infradead.org>]

On Thu, 2009-08-13 at 12:47 +0300, Pekka Enberg wrote:
> Hi Ingo,
> 
> On Thu, 2009-08-13 at 11:36 +0200, Ingo Molnar wrote:
> > * Pekka Enberg <penberg@cs.helsinki.fi> wrote:
> > 
> > > Latest Linus git kernel with -rc4 perf userspace:
> > > 
> > >   penberg@penberg-laptop:~/testing/jato$ perf record -f jato HelloWorldSwing
> > >   [ perf record: Captured and wrote 1.402 MB perf.data (~61261 samples) ]
> > > 
> > >   penberg@penberg-laptop:~/testing/jato$ perf report --sort comm,dso,symbol|head -30
> > >   # Samples: 60457
> > >   #
> > >   # Overhead  Command                            Shared Object  Symbol
> > >   # ........  .......  .......................................  ......
> > >   #
> > > 
> > >   [ Note: the jato executable doesn't appear anywhere in the trace. I'm
> > >     pretty sure the libzip functions, for example, are from jato, not perf.  ]
> > > 
> > >       36.14%     perf  /home/penberg/bin/perf                   [.] dso__synthesize_plt_symbols
> > >       35.20%     perf  /home/penberg/bin/perf                   [.] dso__new
> > >        3.57%     perf  /home/penberg/bin/jato                   [.] 0x0000000000c910
> > 
> > Perhaps it's this commit:
> > 
> > 4d1e00a: perf symbol: Fix symbol parsing in certain cases: use the build-id as a symlink
> > 
> > it does not revert cleanly, so you might want to check out these two 
> > commits:
> > 
> >   4d1e00a
> >   4d1e00a~1
> > 
> > and build perf on both, and check with your anomalous perf.data 
> > (without changing it) whether the output makes most sense.
> 
> The commit seems fine. Like I said, the perf.data file is busted even
> with 2.6.31-rc4 userspace so I think it's a kernel bug.

Hrmm, does a -tip kernel work as expected? I'll go audit the kernel
changes between -rc4 and -rc5.