Re: perf/branch-history branch build broken with NO_DEMANGLE=1 perf hists browser: Print overhead percent value for first-level callchain

From: Arnaldo Carvalho de Melo <acme@kernel.org>
To: Andi Kleen <andi@firstfloor.org>
Cc: Namhyung Kim <namhyung@kernel.org>,
	Peter Zijlstra <a.p.zijlstra@chello.nl>,
	Ingo Molnar <mingo@kernel.org>, Paul Mackerras <paulus@samba.org>,
	Namhyung Kim <namhyung.kim@lge.com>,
	LKML <linux-kernel@vger.kernel.org>, Jiri Olsa <jolsa@redhat.com>,
	David Ahern <dsahern@gmail.com>,
	Frederic Weisbecker <fweisbec@gmail.com>,
	Brendan Gregg <brendan.d.gregg@gmail.com>
Subject: Re: perf/branch-history branch build broken with NO_DEMANGLE=1 perf hists browser: Print overhead percent value for first-level callchain
Date: Mon, 24 Nov 2014 22:17:01 -0300	[thread overview]
Message-ID: <20141125011701.GG13167@kernel.org> (raw)
In-Reply-To: <20141124225244.GS12538@two.firstfloor.org>

Em Mon, Nov 24, 2014 at 11:52:44PM +0100, Andi Kleen escreveu:
> > [root@zoo acme]# perf record -a -g -b sleep 2s
> > [ perf record: Captured and wrote 3.033 MB perf.data (~132504 samples) ]
> > [root@zoo acme]# perf report --stdio --branch-history
> > # To display the perf.data header info, please use --header/--header-only options.
> > BFD: Dwarf Error: Offset (2585882475) greater than or equal to .debug_str size (44321517).
> > BFD: Dwarf Error: Could not find abbrev number 11800.
> > <BIG SNIP>
> > BFD: Dwarf Error: Offset (83496016) greater than or equal to .debug_str size (44321517).
> > BFD: Dwarf Error: Offset (48628447) greater than or equal to .debug_str size (44321517).
> > (END)Segmentation fault (core dumped)
> > [root@zoo acme]# 

> > Will investigate this later today/tomorrow, if nobody finds a fix in the meantime.

> I cannot reproduce this.

> For me it looks like you have some binary or debuginfo that your libbfd
> doesn't like. --branch-history resolves all addresses as srcline, 
> so it will actually walk all the line numbers.

> Can you please find out which one it is? Probably can be seen
> by just going up a few levels in gdb and dumping the event.

> If you can find the address that explodes you can also try it directly with
> addr2line. If that works it's some problem in the perf implementation.

Right, I'll try and figure this out, since you didn't manage to easily
reproduce this.

> It is likely that it would need to be fixed in libbfd.
> 
> To work around it we could turn off force resolving the srcline,
> but that would make the output much less useful too unfortuantely...

I'll try to find out if there is something that can flag a file as
"unsafe" to feed libbfd with, so that we can have a workaround for known
buggy libbfds, if that ends up being really the case.

- Arnaldo