From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <linux-kernel-owner@vger.kernel.org>
Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand
        id S1755309AbdERIho (ORCPT <rfc822;w@1wt.eu>);
        Thu, 18 May 2017 04:37:44 -0400
Received: from mail.kdab.com ([176.9.126.58]:49322 "EHLO mail.kdab.com"
        rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP
        id S1755132AbdERIhk (ORCPT <rfc822;Linux-kernel@vger.kernel.org>);
        Thu, 18 May 2017 04:37:40 -0400
From: Milian Wolff <milian.wolff@kdab.com>
To: Linux-kernel@vger.kernel.org
Cc: linux-perf-users@vger.kernel.org, Milian Wolff <milian.wolff@kdab.com>,
        Arnaldo Carvalho de Melo <acme@redhat.com>,
        David Ahern <dsahern@gmail.com>, Namhyung Kim <namhyung@kernel.org>,
        Peter Zijlstra <a.p.zijlstra@chello.nl>,
        Yao Jin <yao.jin@linux.intel.com>
Subject: [PATCH v3] perf report: fix off-by-one for non-activation frames
Date: Thu, 18 May 2017 10:37:35 +0200
Message-Id: <20170518083735.14126-1-milian.wolff@kdab.com>
Sender: linux-kernel-owner@vger.kernel.org
List-ID: <linux-kernel.vger.kernel.org>
X-Mailing-List: linux-kernel@vger.kernel.org

As the documentation for dwfl_frame_pc says, frames that
are no activation frames need to have their program counter
decremented by one to properly find the function of the caller.

This fixes many cases where perf report currently attributes
the cost to the next line. I.e. I have code like this:

~~~~~~~~~~~~~~~
  #include <thread>
  #include <chrono>

  using namespace std;

  int main()
  {
    this_thread::sleep_for(chrono::milliseconds(1000));
    this_thread::sleep_for(chrono::milliseconds(100));
    this_thread::sleep_for(chrono::milliseconds(10));

    return 0;
  }
~~~~~~~~~~~~~~~

Now compile and record it:

~~~~~~~~~~~~~~~
g++ -std=c++11 -g -O2 test.cpp
echo 1 | sudo tee /proc/sys/kernel/sched_schedstats
perf record \
    --event sched:sched_stat_sleep \
    --event sched:sched_process_exit \
    --event sched:sched_switch --call-graph=dwarf \
    --output perf.data.raw \
    ./a.out
echo 0 | sudo tee /proc/sys/kernel/sched_schedstats
perf inject --sched-stat --input perf.data.raw --output perf.data
~~~~~~~~~~~~~~~

Before this patch, the report clearly shows the off-by-one issue.
Most notably, the last sleep invocation is incorrectly attributed
to the "return 0;" line:

~~~~~~~~~~~~~~~
  Overhead  Source:Line
  ........  ...........

   100.00%  core.c:0
            |
            ---__schedule core.c:0
               schedule
               do_nanosleep hrtimer.c:0
               hrtimer_nanosleep
               sys_nanosleep
               entry_SYSCALL_64_fastpath .tmp_entry_64.o:0
               __nanosleep_nocancel .:0
               std::this_thread::sleep_for<long, std::ratio<1l, 1000l> > thread:323
               |
               |--90.08%--main test.cpp:9
               |          __libc_start_main
               |          _start
               |
               |--9.01%--main test.cpp:10
               |          __libc_start_main
               |          _start
               |
                --0.91%--main test.cpp:13
                          __libc_start_main
                          _start
~~~~~~~~~~~~~~~

With this patch here applied, the issue is fixed. The report becomes
much more usable:

~~~~~~~~~~~~~~~
  Overhead  Source:Line
  ........  ...........

   100.00%  core.c:0
            |
            ---__schedule core.c:0
               schedule
               do_nanosleep hrtimer.c:0
               hrtimer_nanosleep
               sys_nanosleep
               entry_SYSCALL_64_fastpath .tmp_entry_64.o:0
               __nanosleep_nocancel .:0
               std::this_thread::sleep_for<long, std::ratio<1l, 1000l> > thread:323
               |
               |--90.08%--main test.cpp:8
               |          __libc_start_main
               |          _start
               |
               |--9.01%--main test.cpp:9
               |          __libc_start_main
               |          _start
               |
                --0.91%--main test.cpp:10
                          __libc_start_main
                          _start
~~~~~~~~~~~~~~~

Similarly it works for signal frames:

~~~~~~~~~~~~~~~
#include <stdio.h>
#include <stdlib.h>
#include <signal.h>

#define __noinline  __attribute__((noinline))

__noinline void bar(void)
{
  volatile long cnt = 0;

  for (cnt = 0; cnt < 100000000; cnt++);
}

__noinline void foo(void)
{
  bar();
}

void sig_handler(int sig)
{
  foo();
}

int main(void)
{
  signal(SIGUSR1, sig_handler);
  raise(SIGUSR1);

  foo();
  return 0;
}
~~~~~~~~~~~~~~~~

Before, the report wrongly points to `signal.c:29` after raise():

~~~~~~~~~~~~~~~~
$ perf report --stdio --no-children -g srcline -s srcline
...
   100.00%  signal.c:11
            |
            ---bar signal.c:11
               |
               |--50.49%--main signal.c:29
               |          __libc_start_main
               |          _start
               |
                --49.51%--0x33a8f
                          raise .:0
                          main signal.c:29
                          __libc_start_main
                          _start
~~~~~~~~~~~~~~~~

With this patch in, the issue is fixed and we instead get:

~~~~~~~~~~~~~~~~
   100.00%  signal   signal            [.] bar
            |
            ---bar signal.c:11
               |
               |--50.49%--main signal.c:29
               |          __libc_start_main
               |          _start
               |
                --49.51%--0x33a8f
                          raise .:0
                          main signal.c:27
                          __libc_start_main
                          _start
~~~~~~~~~~~~~~~~

Note how this patch fixes this issue for both unwinding methods, i.e.
both dwfl and libunwind. The former case is straight-forward thanks
to dwfl_frame_pc. For libunwind, we replace the functionality via
unw_is_signal_frame for any but the very first frame.

Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Yao Jin <yao.jin@linux.intel.com>
Signed-off-by: Milian Wolff <milian.wolff@kdab.com>
---
 tools/perf/util/unwind-libdw.c           |  6 +++++-
 tools/perf/util/unwind-libunwind-local.c | 11 +++++++++++
 2 files changed, 16 insertions(+), 1 deletion(-)

v2:
- use unw_is_signal_frame to also fix this issue for libunwind

v3:
- only check current signal in libunwind case, which fixes the
  example code provided by Namhyung. The example is added in the
  commit message and I confirmed that libdw and libunwind now
  both produce the correct results for this case

diff --git a/tools/perf/util/unwind-libdw.c b/tools/perf/util/unwind-libdw.c
index f90e11a555b2..943a06291587 100644
--- a/tools/perf/util/unwind-libdw.c
+++ b/tools/perf/util/unwind-libdw.c
@@ -168,12 +168,16 @@ frame_callback(Dwfl_Frame *state, void *arg)
 {
 	struct unwind_info *ui = arg;
 	Dwarf_Addr pc;
+	bool isactivation;
 
-	if (!dwfl_frame_pc(state, &pc, NULL)) {
+	if (!dwfl_frame_pc(state, &pc, &isactivation)) {
 		pr_err("%s", dwfl_errmsg(-1));
 		return DWARF_CB_ABORT;
 	}
 
+	if (!isactivation)
+		--pc;
+
 	return entry(pc, ui) || !(--ui->max_stack) ?
 	       DWARF_CB_ABORT : DWARF_CB_OK;
 }
diff --git a/tools/perf/util/unwind-libunwind-local.c b/tools/perf/util/unwind-libunwind-local.c
index f8455bed6e65..84d553898e2a 100644
--- a/tools/perf/util/unwind-libunwind-local.c
+++ b/tools/perf/util/unwind-libunwind-local.c
@@ -692,6 +692,17 @@ static int get_entries(struct unwind_info *ui, unwind_entry_cb_t cb,
 
 		while (!ret && (unw_step(&c) > 0) && i < max_stack) {
 			unw_get_reg(&c, UNW_REG_IP, &ips[i]);
+
+			/*
+			 * Decrement the IP for any non-activation frames.
+			 * this is required to properly find the srcline
+			 * for caller frames.
+			 * See also the documentation for dwfl_frame_pc,
+			 * which this code tries to replicate.
+			 */
+			if (unw_is_signal_frame(&c) <= 0)
+				--ips[i];
+
 			++i;
 		}
 
-- 
2.13.0