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>
from time import sleep
b = BPF(src_file="task_switch.c")
-b.attach_kprobe(event="finish_task_switch", fn_name="count_sched")
+b.attach_kprobe(event_re="^finish_task_switch$|^finish_task_switch\.isra\.\d$",
+ fn_name="count_sched")
# generate many schedule events
for i in range(0, 100): sleep(0.01)
def setUp(self):
b = BPF(text=text, debug=0)
self.stats = b.get_table("stats")
- b.attach_kprobe(event="finish_task_switch", fn_name="count_sched")
+ b.attach_kprobe(event_re="^finish_task_switch$|^finish_task_switch\.isra\.\d$",
+ fn_name="count_sched")
def test_sched1(self):
for i in range(0, 100):
max_pid = int(open("/proc/sys/kernel/pid_max").read())
b = BPF(text=bpf_text, cflags=["-DMAX_PID=%d" % max_pid])
-b.attach_kprobe(event="finish_task_switch", fn_name="sched_switch")
+b.attach_kprobe(event_re="^finish_task_switch$|^finish_task_switch\.isra\.\d$",
+ fn_name="sched_switch")
print("Tracing %s-CPU time... Hit Ctrl-C to end." %
("off" if args.offcpu else "on"))
# initialize BPF
b = BPF(text=bpf_text)
-b.attach_kprobe(event="finish_task_switch", fn_name="oncpu")
+b.attach_kprobe(event_re="^finish_task_switch$|^finish_task_switch\.isra\.\d$",
+ fn_name="oncpu")
matched = b.num_open_kprobes()
if matched == 0:
print("error: 0 functions traced. Exiting.", file=stderr)
# initialize BPF
b = BPF(text=bpf_text)
-b.attach_kprobe(event="finish_task_switch", fn_name="oncpu")
+b.attach_kprobe(event_re="^finish_task_switch$|^finish_task_switch\.isra\.\d$",
+ fn_name="oncpu")
b.attach_kprobe(event="try_to_wake_up", fn_name="waker")
matched = b.num_open_kprobes()
if matched == 0:
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.")
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 higher than %d us" % min_us)
print("%-8s %-16s %-6s %14s" % ("TIME", "COMM", "TID", "LAT(us)"))