On Thu, 1 Jul 2021 10:38:14 +0200
Jiri Olsa [off-list ref] wrote:

On Thu, Jul 01, 2021 at 08:58:54AM +0900, Masami Hiramatsu wrote:

SNIP

quoted

quoted

quoted

  		return &bpf_override_return_proto;
  #endif
+	case BPF_FUNC_get_func_ip:
+		return &bpf_get_func_ip_proto_kprobe;
  	default:
  		return bpf_tracing_func_proto(func_id, prog);
  	}

diff --git a/kernel/trace/trace_kprobe.c b/kernel/trace/trace_kprobe.c
index ea6178cb5e33..b07d5888db14 100644
--- a/kernel/trace/trace_kprobe.c
+++ b/kernel/trace/trace_kprobe.c

@@ -1570,6 +1570,18 @@ static int kretprobe_event_define_fields(struct trace_event_call *event_call)
  }
  
  #ifdef CONFIG_PERF_EVENTS
+/* Used by bpf get_func_ip helper */
+DEFINE_PER_CPU(u64, current_kprobe_addr) = 0;

Didn't check other architectures. But this should work
for x86 where if nested kprobe happens, the second
kprobe will not call kprobe handlers.

No problem, other architecture also does not call nested kprobes handlers.
However, you don't need this because you can use kprobe_running()
in kprobe context.

kp = kprobe_running();
if (kp)
	return kp->addr;

great, that's easier

quoted

BTW, I'm not sure why don't you use instruction_pointer(regs)?

I tried that but it returns function address + 1,
and I thought that could be different on each arch
and we'd need arch specific code to deal with that

Oh, I got it. Yes, since it emulates the int3 interruption, the
regs->ip must be kp->addr + 1 on x86. And indeed, it depends
on each arch.

Thank you,

thanks,
jirka

-- 
Masami Hiramatsu [off-list ref]