Re: [PATCH 0/2] perf: add sort by inclusive time functionality (v2)

From: Namhyung Kim
Date: Mon Mar 12 2012 - 03:15:49 EST


Hi,

2012-03-09 3:49 AM, Arun Sharma wrote:
On 3/8/12 7:31 AM, Frederic Weisbecker wrote:
On Thu, Mar 08, 2012 at 08:29:01AM +0100, Ingo Molnar wrote:

* Arun Sharma<asharma@xxxxxx> wrote:

This patch series refactors existing code a bit and adds sort by
inclusive time (time spent in the function + callees).

Sample command lines:

# perf record -ag -- sleep 1
# perf report -g graph,0.5,callee -n -s inclusive

So I tried this out with:

$ taskset 1 perf record -g git gc

and got entries above 100% (in the TUI):

$ perf report -g graph,0.5,callee -n -s inclusive

+ 321.11% 5628 [.] 0x392b609269
+ 142.27% 3774 [.] create_delta
+ 78.86% 1248 [.] lookup_object
+ 40.54% 1348 [k] system_call_fastpath
[...]

Is that expected?

Yes - this is the "known bug" I noted in the cover letter

The second column (samples) is still accurate and could be used for the analysis.


I think this happens because of this:

- hists->stats.total_period += h->period;
+ if (!h->inclusive)
+ hists->stats.total_period += h->period;

Which I'm not sure why it is needed btw.

Suppose the perf.data file had 1000 samples each with a period of 1e6 events.
total_period would be 1e9 without -s inclusive. Further, let's say the
callchains had an average length of 10.

Now, after adding extra entries to the histogram, total_period would be 1e10,
which screws up the percentages. I'd like to differentiate between the hist
entries that were in the event stream vs the ones added for inclusive time
computation. Desired end result: the total_period remains unchanged at 1e9.

This is done via:

+ if (i != 0)
+ he->inclusive = 1;
+ else
+ orig_he = he;

Either (i != 0) is not a good enough test, or the inclusive bit is not getting
propagated properly after histogram collapsing/resorting. This is the part I
need to better understand and debug.

I tried to explain this problem in my first RFC message as well:

http://thread.gmane.org/gmane.linux.kernel/1262289

The problem Ingo is running into (and I've reproduced it on my end as well) is
that total_period is smaller than without -s inclusive i.e. h->inclusive is 1
when it shouldn't be.


I think it's because of the shared hist_entry. If a callchain is a subset of another, it will be marked as inclusive so that it cannot be contributed to total period. Say, there're two chains - X (a -> b -> c) and Y (a -> b), once __hists__add_entry_inclusive() was called on X, we have:

a -> b -> c
a -> b (inclusive)
a (inclusive)

And then, calling the function on Y should make:

a -> b
a (inclusive)

However, since both callchains are in tree already they'll be shared and marked *inclusive*. Thus the total period will not increased at all for Y. Also I guess the reverse case - add Y first, and then X - will have the same result.

Thanks,
Namhyung
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/