Re: [PATCH 2/3] perf tools: Add callchain order support for libunwind DWARF unwinder

["Wangnan (F)" <wangnan0@huawei.com>]

On 2015/11/17 23:05, Jiri Olsa wrote:
> From: Jiri Olsa <jolsa@redhat.com>
>
> As reported by Milian, currently for DWARF unwind (both libdw
> and libunwind) we display callchain in callee order only.
>
> Adding the support to follow callchain order setup to libunwind
> DWARF unwinder, so we could get following output for report:
>
>    $ perf record --call-graph dwarf ls
>    ...
>    $ perf report --no-children --stdio
>
>      39.26%  ls       libc-2.21.so      [.] __strcoll_l
>                   |
>                   ---__strcoll_l
>                      mpsort_with_tmp
>                      mpsort_with_tmp
>                      sort_files
>                      main
>                      __libc_start_main
>                      _start
>                      0
>
>    $ perf report -g caller --no-children --stdio
>      ...
>      39.26%  ls       libc-2.21.so      [.] __strcoll_l
>                   |
>                   ---0
>                      _start
>                      __libc_start_main
>                      main
>                      sort_files
>                      mpsort_with_tmp
>                      mpsort_with_tmp
>                      __strcoll_l
>
> Reported-by: Milian Wolff <milian.wolff@kdab.com>
> Based-on-patch-by: Milian Wolff <milian.wolff@kdab.com>
> Link: http://lkml.kernel.org/n/tip-lmtbeqm403f3luw4jkjevsi5@git.kernel.org
> Signed-off-by: Jiri Olsa <jolsa@kernel.org>
> ---
>   tools/perf/util/unwind-libunwind.c | 47 ++++++++++++++++++++++++--------------
>   1 file changed, 30 insertions(+), 17 deletions(-)
>
> diff --git a/tools/perf/util/unwind-libunwind.c b/tools/perf/util/unwind-libunwind.c
> index 0ae8844fe7a6..705e1c19f1ea 100644
> --- a/tools/perf/util/unwind-libunwind.c
> +++ b/tools/perf/util/unwind-libunwind.c

[SNIP]

>   
> -		unw_get_reg(&c, UNW_REG_IP, &ip);
> -		ret = ip ? entry(ip, ui->thread, cb, arg) : 0;

In original code if ip == 0 entry() won't be called.

> +		if (callchain_param.order == ORDER_CALLER)
> +			j = max_stack - i - 1;
> +		ret = entry(ips[j], ui->thread, cb, arg);

But in new code event if ips[j] == 0 an entry will be built, which causes
a behavior changes user noticable:

Before this patch:


# perf report --no-children --stdio --call-graph=callee
...
      3.38%  a.out    a.out             [.] funcc
               |
               ---funcc
                  |
                   --2.70%-- funcb
                             funca
                             main
                             __libc_start_main
                             _start

After this patch:

# perf report --no-children --stdio --call-graph=callee
...
      3.38%  a.out    a.out             [.] funcc
               |
               ---funcc
                  |
                  |--2.70%-- funcb
                  |          funca
                  |          main
                  |          __libc_start_main
                  |          _start
                  |
                   --0.68%-- 0


I'm not sure whether we can regard this behavior changing as a bugfix? I 
think
there may be some reason the original code explicitly avoid creating an '0'
entry.

Then I tried to find why perf can't get call frame on my case, and
I guess there's something wrong whe dealing with 'call' command, because
the instruction on it I can't get callchain from libunwind is a 'callq':

   ...
   4005bf:       be 00 00 00 00          mov    $0x0,%esi
   4005c4:       48 89 c7                mov    %rax,%rdi
   4005c7:       e8 74 fe ff ff          callq  400440 <gettimeofday@plt>
                 us2 = tv2.tv_sec * 1000000 + tv2.tv_usec;
   4005cc:       48 8b 04 24             mov    (%rsp),%rax
   ...

But this is another problem, we can discuss it in a new thread.

Thank you.