What is the different(and stand for) for cycles:Pu(self) and cycles:Pu(incl.) in the caller/callee panel?

[zw963]

Sorry for ask question here, because i am a newbie for use perf profile, and i search on all search engine for find it out, but no any clue.

I generate profile use following command:

$: perf record --call-graph dwarf -- bin/tartrazine ../crycco/src/crycco.cr swapoff

Then, use hotspot open the perf.data, like following:

$: hotspot perf.data

I get a window like following screenshot.

I don't know what cycles:Pu(self) and cycles:Pu(incl.) stand for in above caller/callee panel, sure i can't understand why others functions those two column value almost same, but, for char_bytesize_at function, why self is 8.83%, and incl. is 88.9%?

is there a link for keyword for study knowledge of this? could you plese provide some guidance? Thank you.
Thanks

[milianw]

self mean samples that point directly at code from this function, and incl. means cost of this function including all other functions called by it.

the :P modifier is the "as precise as possible" modifier for x86/intel PMU - I think the kernel documentations will talk about this somwhere, maybe in man perf record. the :u modifier means "only userspace" - i.e. excluding cost in kernel space.

does this answer your questions?

[zw963]

man perf record indeed many useful information, I will definitely read it thoroughly.

Thanks a lot.

[zw963]

One more question, is there a way to get a function invoke count?

e.g. above char_bytesize_at, i want know this function took a lot of time indeed when invoke once, or it has been called very many times.

It that possible for hotspot?

[milianw]

No, a sampling profiler does not know what happens inbetween samples - it could be many repeated function calls or just one long running one. But you can use trace points for this purpose, e.g. via perf probe (also see its manpage on uprobe)