[lkml]   [2011]   [Mar]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH] Add inverted call graph report support to perf tool
On Thu, Mar 10, 2011 at 10:32:43PM +0800, Sam Liao wrote:
> On Thu, Mar 10, 2011 at 10:43 AM, Frederic Weisbecker
> <> wrote:
> > On Tue, Mar 08, 2011 at 04:59:30PM +0800, Sam Liao wrote:
> >> On Tue, Mar 8, 2011 at 2:06 AM, Frederic Weisbecker <> wrote:
> >> > So, instead of having such temporary copy, could you rather feed the callchain
> >> > into the cursor in reverse from perf_session__resolve_callchain() ?
> >> >
> >> > You can keep the common part inside the loop into a seperate helper
> >> > but have two different kinds of loops.
> >>
> >> In perf_session__resolve_callchain, only the callchain itself can be reversed,
> >> which means the root of report will still be the ip of the event with a reversed
> >> call chain sub tree. But what is more impressive to user is to make "main" like
> >> function to be the root of the report, and this means that both the ip
> >> and call chain is
> >> involved to the reversion process.
> >>
> >> Since the ip of event is resolved in event__preprocess_sample, so it is kind
> >> hard to do such reversion in a better way.
> >
> > You are making an interesting point.
> >
> > My view of this feature was limited to the current per hist area: having
> > the callchains on top of hists that can be sorted per ip, dso, pid, etc...
> > like we have today basically. So my view was for this reverse callchain
> > to show us one callers profiling for each hist entry.
> >
> > But your idea of turning the callee into the caller would show us a very global
> > profiling. With reverse callchains it can be a very nice overview of the big picture.
> >
> > IMO both workflow can be interesting:
> >
> > 1) Have a big reversed callchain overview, with one root per entrypoint. This
> > what you wanted.
> > 2) Have a per hist 1)  which means a per hist per entrypoint callchain
> >
> > 1) involves reverting both callchains and ip <->caller whereas 2) only involves
> > reverting the callchain.
> Having both workflow included would be more helpful.

That's the point, we should be able to do both. But only 1) is possible with
your initial proposition.

> >
> > In order to get both features with a maximum flexibility and keep that extendable, I
> > would suggest to decouple that in two independant parts:
> >
> >        - an option to get reversed callchains. Using the -g option and caller/callee
> >        as a third argument.
> >
> This could be easily extended by reversing the callchain symbols as
> you mentioned.

Yeah. -g caller only requires to iterate the callchain in reverse.

> >        - a new "caller" sort entry. What defines a hist entry is a set of sort
> >        entries: dso, symbol, pid, comm, ... That we use with the -s option in perf report.
> >        If you want one hist per entrypoint, we could add a new "caller" sort entry.
> >        Then perf report -s caller will (roughly) produce one hist for main(), one hist
> >        for kernel_thread(), etc...
> >
> I'm not sure adding a "caller" sort entry can get things done. As for
> my limited understanding,
> "sort" is kind way to group events

This is actually _what_ group events. This defines how hist entries are

If you do "perf report -s sym", events will be grouped by symbols.
Thus if you had thousands events but all of them only hit sym1 and sym2
then you'll see two groups in your histogram.


# ./perf report -s sym --stdio
# Events: 4 cycles
# Overhead Symbol
# ........ .................
36.72% [.] hex2u64
31.21% [k] __lock_acquire
18.03% [k] lock_acquire
14.04% [k] sub_preempt_count

We may have got thousand events for the above profile. But only 4 symbols
were hit in amongst these thousand events. As we asked for, events have been
grouped per symbol target.

Callchains follow this grouping scheme. Below the __lock_acquire hist,
you would only get callchains for which the root (deepest callee) was __lock_acquire.

If you have several grouping, like -s sym, dso, pid
then it computes an intersection. Events will be grouped when their
sym, dso and pid are equal. Moreoever they will be sorted, first dimension
per sym, second dimension per dso, third dimension per pid.

You should play a bit with different combinations to get the whole picture
and how it works.

Callchains still follow the grouping, as elaborated as it can be. For the hist
that has sym1, dso2 and pid 3, you'll find only callchains that start from sym1
for events that happened on dso2 and pid3.

, after we group all the events
> under "main" or "kernel_thread",
> the sub-trees will still rooted as ip entry points with a reversed
> call-chain sub-trees which seems
> just the same as the previous workflow. Am I right? If so, here we
> still have to revert the ip and
> callchain.

No. The callchain will follow that grouping. If you group only per caller
(-s caller) you may have one hist entry for main and another for kernel_thread.
Then below the main entry, you'll have only callchains starting
from main. And below the kernel_thread, only callchains starting from kernel_thread.

It depends if you select reverse callchain or not:

$ perf report -s caller

That will report main and kernel_thread as hists, and regular callee -> caller callchains.
Hence under main hist, you'll a lot of callchain starting from random points and all
ending in main!

$ perf report -s caller -g caller

That will report main and kernel_thread as hists, with callchains starting from
main under main.

It becomes interesting when you want more granularity with -s caller,dso if we bring a way
to push forward the entrypoint one day. I suspect even more sorting combinations are
going to be interesting.

> > Hence, someone running "perf report -g fractal,0.5,reversed -s dso" is going to have a
> > per dso caller-callchain profiling.
> >
> > Someone running "perf report -g fractal,0.5,reversed -s caller" is going to have that
> > global caller profiling you wanted. You'll have one hist per entrypoint.
> >
> > The caller sorting mode may sound a bit limited currently, but think about what happens
> > when you push forward the entrypoint, if one day we bring a feature to filter the callchains
> > on some dso address space , we could do a caller callchain profile starting on a shared library
> > and pinpoint which functions are mostly called on it, so that can be coupled with dso sorting mode,
> > etc... So that looks like a right way to go.
> This is interesting, we might choose different caller based on dso
> instead of the process/thread entry
> point.

Exactly! That's for later, but cutting the things like I'm suggesting
would pave the way for that.

> This also remind me one thing, at present, though with
> different sort orders, the event trees still
> root in the ip of the event.

Currently yeah.

> For an event with call chain like:
> main->func1->func2->lib_func3->ip, what
> about we expand this event to following events:
> - a direct event with ip called, with full call chain.
> - a indirect event with libfunc3 called, with lib_func3->ip call chain
> - a indirect event with func2 called ...
> - a indirect event with main called,
> With these indirect event added(we may also need to add direct count
> and indirect count), we are
> able to check any routine in any dso which functions are mostly
> called.

I see your point, and it may be a good idea in theory but I fear
the O(hell) (sorry I suck in math but I figure out the pain)
property is going to prevent this feature from ever being usable.
Regular callchains can already be slow to process on of
hundreds MB. With such a feature their processing will simply never

I suspect we should rather opt for fileting callchain dso
domains. Like only start from a given dso or so.

> the sysprof that Ingo mentioned.

sysprof does something like that?

Note, another reason to avoid abusing the event.ip to group
per caller, is that we could be able to:

perf report -s caller,sym -g caller

If we limit callchains to start from a given, this may
sort hists per caller and then per endpoint.

If your library offers function func1, func2, etc... It will sort
them per usage (func1 has been first used, then func2, etc...)
then per endpoint overhead (func1 most often sticks in strcpy(),
then on read(), etc....).


That may or may not be useful. I don't know. In fact I don't
want to take the responsibility to judge whether it's useful
or not. Thus I prefer caller and ip to be two different
properties of hist entries and not having one absusing the
other, so that we don't prevent this feature to exist (or many
other sort combinations I haven't imagined).
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

 \ /
  Last update: 2011-03-11 12:59    [W:0.075 / U:4.792 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site