perf report: Fix off-by-one for non-activation frames
authorMilian Wolff <milian.wolff@kdab.com>
Wed, 24 May 2017 06:21:25 +0000 (15:21 +0900)
committerIngo Molnar <mingo@kernel.org>
Wed, 24 May 2017 06:41:48 +0000 (08:41 +0200)
commit1982ad48fc82c284a5cc55697a012d3357e84d01
tree9d8564f721fb767c8b55bd80d33d03af343a9a5f
parentb21cc97810932a551f7aac46f0b89c469c828b3f
perf report: Fix off-by-one for non-activation frames

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:

~~~~~~~~~~~~~~~
  __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.

Signed-off-by: Milian Wolff <milian.wolff@kdab.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Yao Jin <yao.jin@linux.intel.com>
Cc: kernel-team@lge.com
Link: http://lkml.kernel.org/r/20170524062129.32529-4-namhyung@kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
tools/perf/util/unwind-libdw.c
tools/perf/util/unwind-libunwind-local.c