Re: [PATCH 0/1] perf report: append inlines to non-dwarf callchains

[Adrian Hunter <adrian.hunter@intel.com>]

On 4/04/23 08:47, Namhyung Kim wrote:
> On Mon, Apr 3, 2023 at 1:30 PM Arnaldo Carvalho de Melo <acme@kernel.org> wrote:
>>
>> Em Fri, Mar 31, 2023 at 10:52:24AM +0200, Artem Savkov escreveu:
>>> On Thu, Mar 30, 2023 at 08:06:20AM +0300, Adrian Hunter wrote:
>>>> On 22/03/23 21:44, Arnaldo Carvalho de Melo wrote:
>>>>> Em Wed, Mar 22, 2023 at 11:18:49AM -0700, Namhyung Kim escreveu:
>>>>>> On Fri, Mar 17, 2023 at 12:41 AM Artem Savkov <asavkov@redhat.com> wrote:
>>>>>>>
>>>>>>> On Thu, Mar 16, 2023 at 02:26:18PM -0700, Namhyung Kim wrote:
>>>>>>>> Hello,
>>>>>>>>
>>>>>>>> On Thu, Mar 16, 2023 at 6:36 AM Artem Savkov <asavkov@redhat.com> wrote:
>>>>>>>>>
>>>>>>>>> In an email to Arnaldo Andrii Nakryiko suggested that perf can get
>>>>>>>>> information about inlined functions from dwarf when available and then
>>>>>>>>> add it to userspace stacktraces even in framepointer or lbr mode.
>>>>>>>>> Looking closer at perf it turned out all required bits and pieces are
>>>>>>>>> already there and inline information can be easily added to both
>>>>>>>>> framepointer and lbr callchains by adding an append_inlines() call to
>>>>>>>>> add_callchain_ip().
>>>>>>>>
>>>>>>>> Looks great!  Have you checked it with perf report -g callee ?
>>>>>>>> I'm not sure the ordering of inlined functions is maintained
>>>>>>>> properly.  Maybe you can use --no-children too to simplify
>>>>>>>> the output.
>>>>>>>
>>>>>>> Thanks for the suggestion. I actually have another test program with
>>>>>>> functions being numbered rather than (creatively) named, so it might be
>>>>>>> easier to use it to figure out ordering. Here's the code:
>>>>>>
>>>>>> Yep, looks good.
>>>>>>
>>>>>> Acked-by: Namhyung Kim <namhyung@kernel.org>
>>>>>
>>>>> So, I'll apply this shorter patch instead, ok?
>>>>>
>>>>> - Arnaldo
>>>>>
>>>>> diff --git a/tools/perf/util/machine.c b/tools/perf/util/machine.c
>>>>> index 803c9d1803dd26ef..abf6167f28217fe6 100644
>>>>> --- a/tools/perf/util/machine.c
>>>>> +++ b/tools/perf/util/machine.c
>>>>> @@ -44,6 +44,7 @@
>>>>>  #include <linux/zalloc.h>
>>>>>
>>>>>  static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock);
>>>>> +static int append_inlines(struct callchain_cursor *cursor, struct map_symbol *ms, u64 ip);
>>>>>
>>>>>  static struct dso *machine__kernel_dso(struct machine *machine)
>>>>>  {
>>>>> @@ -2322,6 +2323,10 @@ static int add_callchain_ip(struct thread *thread,
>>>>>   ms.maps = al.maps;
>>>>>   ms.map = al.map;
>>>>>   ms.sym = al.sym;
>>>>> +
>>>>> + if (append_inlines(cursor, &ms, ip) == 0)
>>>>> +         return 0;
>>>>> +
>>>>>   srcline = callchain_srcline(&ms, al.addr);
>>>>>   return callchain_cursor_append(cursor, ip, &ms,
>>>>>                                  branch, flags, nr_loop_iter,
>>>>
>>>> This seems to be breaking --branch-history.  I am not sure
>>>> append_inlines() makes sense for branches.  Maybe this should be:
>>>>
>>>>     if (!branch && !append_inlines(cursor, &ms, ip))
>>>>             return 0;
>>>>
>>>
>>> Right. So when cllchain_cursor is appended through append_inlines it
>>> always discards branch information, even for the non-inlined function.
>>> So adding !branch makes sense to me. Does anyone else see any problems
>>> with that?
>>
>> I'm no expert in this specific area, so for now till we get to a
>> conclusion on this, I'll follow Andi's suggestion and revert this patch.
> 
> I think we can simply apply Adrian's patch above.

Yes.  The thing is, inserting inline functions into a branch
stack doesn't work very well.  For example, there can be multiple
branches in the same inline function, so adding the inlines on
every branch makes a mess.  Then there is how to handle branching
out of, and then back into, the same inline function - or is it
a second instance of the inline function...  All too hard, so just
don't try.  At least for now.