Message-ID: <20170524150237.GA27760@danjae.aot.lge.com>
Date:   Thu, 25 May 2017 00:02:37 +0900
From:   Namhyung Kim <namhyung@...nel.org>
To:     Milian Wolff <milian.wolff@...b.com>
Cc:     Linux-kernel@...r.kernel.org, linux-perf-users@...r.kernel.org,
        kernel-team@....com
Subject: Re: [PATCH 0/7] generate full callchain cursor entries for inlined
 frames

On Wed, May 24, 2017 at 03:42:59PM +0200, Milian Wolff wrote:
> On Wednesday, May 24, 2017 1:46:04 PM CEST Milian Wolff wrote:
> > On Monday, May 22, 2017 11:06:43 AM CEST Namhyung Kim wrote:
> > > Why not making the fake symbol has start addr of the sample IP and
> > > length of 1.  The histogram sort code also compares the sym->start
> > > which might confuse the output of the children mode too IMHO.
> > 
> > I can try that out, thank you for the suggestion. But I think it can easily
> > break in different ways. I.e. when the same inline function gets used at
> > different IPs, it should actually be considered to be the same function when
> > we group/merge/aggregate. I updated the `match_chain` function accordingly,
> > to do a symname / srcline comparison on inlined frames, instead of relying
> > on the symbol start/end. I think using the IP for the fake symbols won't be
> > more reliable here, don't you think?
> > 
> > In the end, I think we'll always have to special-case inlined fake symbols
> > when we aggregate data, since the sym start/end is always going to be some
> > arbitrary value that may or may not be what we want it to be. Doing the
> > explicit comparison on e.g. srcline/symname is always going to be the most
> > reliable option, as it also directly results in a proper aggregation based
> > on the strings that the user will see in the end.
> 
> I haven't yet tried it out, but I think I can come up with a way to break your 
> approach easily. Assume the following pseudo-code:
> 
> void tail()
> {
>     instr1; // IP1
>     instr2; // IP2
> }
> 
> void mid()
> {
>     tail();
> }
> 
> void main()
> {
>     mid();
> }
> 
> Now, assume both `tail` and `mid` get inlined into `main`. If we get one 
> sample each for both IP1 and IP2, we want the following merged structure if we 
> merge based on symbol:
> 
> sym  | incl | self
> main | 2    | 0
> mid  | 2    | 0
> tail | 2    | 2
> 
> If we would give the inlined fake-symbols a start of the IP, i.e. either IP1 
> or IP2, then we would end up with this (unexpected) behavior instead:
> 
> sym  | incl | self
> main | 2    | 0
> mid  | 1    | 0
> mid  | 1    | 0
> tail | 1    | 1
> tail | 1    | 1
> 
> The reason is that the fake symbols for the inlined frames would be considered 
> to be different functions since their start/end are not equal. This is "wrong" 
> in my eyes - we really have to do symbol name comparisons for inlined frames, 
> and also include srcline if that is desired.

That would depend on how we treat inlined function instances.  Each
instance might be considered as same or not - but I think it'd be
better treating them as same for simplicity.

Also currently perf aggregates samples using symbol name, but a new
sort key might be added to use symbol address later.  Thus it'd be
better to be prepared for such change.

> 
> If you think the above is not a valid assessment, I'll try to change my patch 
> series to use the IP + 1 trick you suggest. But I really don't think it's 
> going to work.

So I agree that we should do symbol name comparison, but I still
prefer setting fake symbol address to [IP, +1].  That would reduce
memory space for annotate as well.

Thanks,
Namhyung


> 
> Cheers
> -- 
> Milian Wolff | milian.wolff@...b.com | Software Engineer
> KDAB (Deutschland) GmbH&Co KG, a KDAB Group company
> Tel: +49-30-521325470
> KDAB - The Qt Experts

Powered by blists - more mailing lists