In GCC10.2 suffix '.isra.0' was appended to 'finish_task_switch'
When buildiing kernel with GCC10 [2] in Debian on an Arm64 machine, it was
found the new "inter-procedural optimization improvements" [1] makes symbol
name 'finish_task_switch' changed to 'finish_task_switch.isra.0'.
Details:
The results, when built with gcc 9.3:
nm ../linux.buildout/kernel/sched/core.o | grep finish_task_switch
0000000000001288 t finish_task_switch
However, when built with gcc 10.2:
nm ../linux.buildout/kernel/sched/core.o | grep finish_task_switch
00000000000012d0 t finish_task_switch.isra.0
The same symbols (with xxx.isra.0 or without, respectively of course) also
appear in final file 'System.map' and in '/proc/kallsyms' when booting. This
negatively impact the tracing tools commonly used in kernel debugging, such
as bcc tools offcputime and runqlat. They hardcode 'finish_task_switch'
(without the .isra.0 suffix) into their scripts.
This patch fix the issue by changing the hardcoded 'finish_task_switch' string
to a python regular expression pattern who can match both the traditional form
'finish_task_switch' and the new gcc10 form 'finish_task_switch.isra.0'
(with single digit at the end). attach_kprobe()'s input parameter 'event_re'
is used for this type of pattern matching.
[1] https://gcc.gnu.org/gcc-10/changes.html
[2] ARCH=arm64 make Image
Signed-off-by: Guodong Xu <guodong.xu@linaro.org>
diff --git a/tools/runqlat.py b/tools/runqlat.py
index 9fd4064..b13ff2d 100755
--- a/tools/runqlat.py
+++ b/tools/runqlat.py
@@ -252,7 +252,8 @@
if not is_support_raw_tp:
b.attach_kprobe(event="ttwu_do_wakeup", fn_name="trace_ttwu_do_wakeup")
b.attach_kprobe(event="wake_up_new_task", fn_name="trace_wake_up_new_task")
- b.attach_kprobe(event="finish_task_switch", fn_name="trace_run")
+ b.attach_kprobe(event_re="^finish_task_switch$|^finish_task_switch\.isra\.\d$",
+ fn_name="trace_run")
print("Tracing run queue latency... Hit Ctrl-C to end.")