Perf Tool Linux
I am tring to use perf version 3.0.8
to profile my embedded system Linux kernel version 3.0.8
.
I cross-compiled tools/perf using zlib and elfutils.
In order to familiarize with this tool I am using a simple quicksort example compiled using -g
option.
I run perf record -F 99 -g -- ./qsort
(copied from the example) and then perf report --stdio
to see my profile results.
This is a snapshot of the perf report -g --stdio
output:
No kallsyms or vmlinux with build-id f41dfb5606beda1d75171db69f1e9466cb4868ba was found
[hi3518_dsu] with build id f41dfb5606beda1d75171db69f1e9466cb4868ba not found, continuing without symbols
No kallsyms or vmlinux with build-id a0f9a05faf6bcfb72282a03e048b47e8a8c0f2ec was found
[hi3518_tde] with build id a0f9a05faf6bcfb72282a03e048b47e8a8c0f2ec not found, continuing without symbols
# Events: 256 cpu-clock
#
# Overhead Command Shared Object Symbol
# ........ ....... ................. ......................
#
96.88% qsort qsort [.] partition
0.78% qsort [kernel.kallsyms] [k] __do_softirq
0.39% qsort [kernel.kallsyms] [k] __wake_up_sync_key
0.39% qsort [kernel.kallsyms] [k] get_page_from_freelist
0.39% qsort [kernel.kallsyms] [k] __memzero
0.39% qsort [kernel.kallsyms] [k] div_s64_rem
0.39% qsort [hi3518_tde] [k] 0x58
0.39% qsort [hi3518_dsu] [k] 0x8154
As you can see there is not call-trace at all. I am really interested to study in deep my system and without a call-trace these results are useless.
You might have some luck using the -fno-omit-frame-pointer
flag while compiling the program you are trying to profile. I am also having issues making call graphs work with perf on an ARM embedded system.
I found a solution. Run perf --call-graph dwarf -- ./qsort
Make sure you have compiled the program with dwarf debug symbols (-g, -ggdb, or -gdwarf flag). In addition, you might need to install debug symbols for any libraries you are using on your system. https://fedoraproject.org/wiki/StackTraces shows how to install debug symbol packages for Fedora.
If the perf command fails with --call-graph dwarf, you may need to compile perf with libunwind in addition to elfutils.
Probably my pref version is really old (or/and the kernel tool). I tried with another soc (with a newer kernel and perf tool) and the callstack works properly. However, even without this feature, I'm using this tool profitably.
链接地址: http://www.djcxy.com/p/80040.html上一篇: 用Babeltrace建筑Perf(Perf到CTF转换)
下一篇: Perf工具Linux