On Mon, Jan 9, 2023 at 5:56 PM Eric Dumazet [off-list ref] wrote:

Note: At least for CONFIG_RETPOLINE=y and gcc 12.2, compiler adds many
additional instructions (and additional memory reads),
even when the trace point is not enabled.

Contrary to some belief, adding a tracepoint is not always 'free'.
tail calls for example are replaced with normal calls.

        .popsection

# 0 "" 2
#NO_APP
.L106:
# net/socket.c:1008: }
        movl    %ebx, %eax      # <retval>,
        popq    %rbx    #
        popq    %rbp    #
        popq    %r12    #
        ret
.L111:
# ./include/trace/events/sock.h:308: DEFINE_EVENT(sock_msg_length,
sock_recv_length,

Hi Eric,  Thanks for your reply,  In fact, it is because the
definition of the tracepoint function is inline,
Not just these two tracepoints，right?

#define __DECLARE_TRACE(name, proto, args, cond, data_proto)            \
      ...
      static inline void trace_##name(proto)

Regarding the above issue, I plan to optimize it like this:

static noinline void call_trace_sock_send_length(struct sock *sk, __u16 family,
                                            __u16 protocol, int ret, int flags)
{
        trace_sock_send_length(sk, family, protocol, ret, 0);
}

static inline int sock_sendmsg_nosec(struct socket *sock, struct msghdr *msg)
{
        int ret = INDIRECT_CALL_INET(sock->ops->sendmsg, inet6_sendmsg,
                                     inet_sendmsg, sock, msg,
                                     msg_data_left(msg));
        BUG_ON(ret == -EIOCBQUEUED);

        if (trace_sock_send_length_enabled()) {
                call_trace_sock_send_length(sock->sk, sock->sk->sk_family,
                                            sock->sk->sk_protocol, ret, 0);
        }
        return ret;
}

What do you think?

Thanks,
Yunhui