[RFC/PATCHSET 0/3] perf tools: Callchain improvement with --children and -g caller
From: Namhyung Kim
Date: Thu Aug 14 2014 - 02:01:47 EST
Hello,
This patchset tries to improve the callchain behavior when caller
ordering (which is not default) and --children (which is default)
enabled. I think that we might need to change default to caller
ordering when --children is enabled - but it needs more discussion of
course.
The "perf report -g" displays callchains callee-first order by default.
That means it can see the callers of the sample in a reverse order.
For example, "intel_idle" entry shows following callchain on my data.
Children Self Command Shared Object Symbol
-------- -------- --------------- ------------------ ----------------
40.53% 40.53% swapper [kernel.kallsyms] [k] intel_idle
|
--- intel_idle
cpuidle_enter
cpuidle_wrap_enter
cpuidle_enter_tk
cpuidle_idle_call
cpu_idle
|
|--85.25%-- start_secondary
|
--14.75%-- rest_init
start_kernel
x86_64_start_reservations
x86_64_start_kernel
So "intel_idle" was called by "cpuidle_enter", and in turn, it's
called by "cpuidle_wrap_enter" and so on.
When with -g "caller", it shows callchains in a reversed order -
"caller-first". I know it's sometimes useful but never used it
by myself though.
40.53% 40.53% swapper [kernel.kallsyms] [k] intel_idle
|
|--85.25%-- start_secondary
| cpu_idle
| cpuidle_idle_call
| cpuidle_enter_tk
| cpuidle_wrap_enter
| cpuidle_enter
| intel_idle
|
--14.75%-- x86_64_start_kernel
x86_64_start_reservations
start_kernel
rest_init
cpu_idle
cpuidle_idle_call
cpuidle_enter_tk
cpuidle_wrap_enter
cpuidle_enter
intel_idle
However, with --children feature added, it now can show all *callees*
of the entry and I think it could be very useful. For example,
"start_kernel" entry now can display it calls rest_init and in turn
cpu_idle and then cpuidle_idle_call (95.72%).
6.14% 0.00% swapper [kernel.kallsyms] [k] start_kernel
|
--- start_kernel
rest_init
cpu_idle
|
|--97.52%-- cpuidle_idle_call
| cpuidle_enter_tk
| |
| |--99.91%-- cpuidle_wrap_enter
| | cpuidle_enter
| | intel_idle
| --0.09%-- [...]
--2.48%-- [...]
Note that start_kernel has no self overhead - meaning that it never
get sampled by itself but constructs such a nice callgraph. But,
sadly, if an entry has self overhead, callchain will get confused with
generated callchain (like above) and self callchains (which reversed
order) like the eariler example.
To be consistent with other entries, I'd like to make it just to show
a single entry - itself - like below since it doesn't have callees
(children) at all. But still use the whole callchain to construct
children entries (like the start_kernel) as usual.
40.53% 40.53% swapper [kernel.kallsyms] [k] intel_idle
|
--- intel_idle
Note that this is just for --children is enabled. Without that, it
will do the original behavior.
Okay, here is a real case other that pure self (intel_idle) and pure
generated (start_kernel) entries. Without this patch the "cpu_idle"
entry in the above data shows callchain info like following. Note
that it has 0.15% of self overhead.
$ perf report --children --g caller
...
# Children Self Command Shared Object Symbol
# ........ ........ ............... .................. .................
#
53.42% 0.15% swapper [kernel.kallsyms] [k] cpu_idle
|--99.72%-- cpu_idle
| |
| |--84.44%-- cpuidle_idle_call
| | |
| | |--92.67%-- cpuidle_enter_tk
| | |
| | /* snip */
| |
| |--7.83%-- tick_nohz_idle_exit
| | |
| | |--77.61%-- hrtimer_cancel
| | | hrtimer_try_to_cancel
| | | __remove_hrtimer
| | | timerqueue_del
| | | rb_erase
| | |
| | --22.39%-- account_idle_ticks
| | irqtime_account_process_tick.isra.2
| |
| |--7.07%-- schedule
| | __schedule
| | perf_event_task_sched_out
| |
| --0.65%-- cpuidle_get_cpu_driver
|
--0.28%-- x86_64_start_kernel /* <==== XXX */
x86_64_start_reservations
start_kernel
rest_init
cpu_idle
As you can see, mostly it just calls other functions - cpuidle_idle_call,
tick_nohz_idle_exit schedule or cpuidle_get_cpu_driver. But 0.28% of it
shows as it it called x86_64_start_kernel which is not correct. It just
came from the reversed self callchain info.
With this patchset applied, the incorrect 0.28% is disappeared and it
moves under cpu_idle itself (with [...]).
53.42% 0.15% swapper [kernel.kallsyms] [k] cpu_idle
|
--- cpu_idle
|
|--84.20%-- cpuidle_idle_call
| |
| |--92.67%-- cpuidle_enter_tk
| | |
| | /* snip */
|
|--7.81%-- tick_nohz_idle_exit
| |
| |--77.61%-- hrtimer_cancel
| | hrtimer_try_to_cancel
| | __remove_hrtimer
| | timerqueue_del
| | rb_erase
| |
| --22.39%-- account_idle_ticks
| irqtime_account_process_tick.isra.2
|
|--7.05%-- schedule
| __schedule
| perf_event_task_sched_out
|
|--0.65%-- cpuidle_get_cpu_driver
--0.28%-- [...] /* <===== this is the self overhead */
You can get this from 'perf/caller-chain-v1' branch on my tree
git://git.kernel.org/pub/scm/linux/kernel/git/namhyung/linux-perf.git
Any comments are welcomed, thanks
Namhyung
Namhyung Kim (3):
perf report: Relax -g option parsing not to limit the option order
perf tools: Put callers above callee when callchain order is caller
perf callchain: Prune misleading callchains for self entries
tools/perf/ui/hist.c | 8 ++--
tools/perf/util/callchain.c | 95 ++++++++++++++++++---------------------------
tools/perf/util/hist.c | 23 ++++++++++-
3 files changed, 65 insertions(+), 61 deletions(-)
--
2.0.0
--
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/