* Jiri Olsa [off-list ref] [2012-01-16 16:33:27]:

On Mon, Jan 16, 2012 at 08:15:38PM +0530, Srikar Dronamraju wrote:

quoted

quoted

I've tested following event:
        echo "p:probe_libc/free /lib64/libc-2.13.so:0x7a4f0 %ax" > ./uprobe_events

and commands like:
        perf record -a -e probe_libc:free  --filter "common_pid == 1127"
        perf record -e probe_libc:free --filter "arg1 == 0xa" ls

got me proper results.

Okay thanks for the inputs.

quoted

thanks,
jirka

---
The preemption needs to be disabled when submitting data into perf.

I actually looked at other places where perf_trace_buf_prepare and
perf_trace_buf_submit are being called. for example perf_syscall_enter
and perf_syscall_exit both call the above routines and they didnt seem
to be called with premption disabled. Is that the way perf probe is
called in our case that needs us to call pre-emption here? Did you see a
case where calling these without preemption disabled caused a problem?

the perf_trace_buf_prepare touches per cpu variables,
hence the preemption disabling

the perf_trace_buf_prepare code is used by syscalls,
kprobes, and trace events

- both syscalls and trace events are implemented by
  tracepoints which disable preemption before calling the probe
  (see __DO_TRACE macro in include/linux/tracepoint.h)

- kprobes disable preemption as well
  (kprobe_handler in arch/x86/kernel/kprobes.c)
  haven't checked the optimalized kprobes,
  but should be the same case

jirka

Okay, thanks for the clarification, I will pick the fix.

--
Thanks and Regards
Srikar

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>