[PATCH 3/3] perf: parse the .debug_frame section in case .eh_frame is not present
Jean Pihet
jean.pihet at newoldbits.com
Thu Sep 5 09:05:57 EDT 2013
Hi Will,
On Thu, Sep 5, 2013 at 2:45 PM, Will Deacon <will.deacon at arm.com> wrote:
> Hi Jean,
>
> [adding Michael, since I know he was interested in this]
>
> On Wed, Sep 04, 2013 at 07:04:14PM +0100, Jean Pihet wrote:
>> On ARM the debug info is not present in the .eh_frame sections but
>> instead in .debug_frame.
>> Use libunwind to load and parse the debug info.
>
> How have you tested this? Regardless of whether or not I apply this patch, I
> get the same (broken/truncated) callchains for userspace in perf report.
Here are the commands I have been using:
perf record -g dwarf -- <binary to profile>
perf report --sort symbol --call-graph --stdio
>
> E.g. the following stupid program (built with -O0 -g):
>
> --->8
>
> void bar(void)
> {
> int i;
> for (i = 0; i < 1000000; ++i)
> asm volatile("nop" ::: "memory");
> }
>
> void foo(void)
> {
> bar();
> }
>
>
> int main(void)
> {
> foo();
> return 0;
> }
>
> 8<---
>
> Gives me an incomplete callchain:
>
> # Overhead Command Shared Object Symbol
> # ........ ........ ................. ...............................
> #
> 0.00% unwindme unwindme [.] bar
> |
> --- bar
I get the following with a simple stupid program with a long call chain:
0.57% stress_bt stress_bt [.] foo_115
|
--- foo_115
foo_114
foo_113
...
foo_92
bar
doit
main
__libc_start_main
Things to check:
- compile the binaries and libraries with -g (-dbg flavor of libs are
usually ok),
- use -g dwarf in perf record
> This is the same with or without your patch.
>
> Will
Thanks for testing!
Jean
>
> _______________________________________________
> linux-arm-kernel mailing list
> linux-arm-kernel at lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
More information about the linux-arm-kernel
mailing list