Re: [External] Re: [PATCH v4] sock: add tracepoint for send recv length

[Steven Rostedt]

On Mon, 9 Jan 2023 15:54:38 +0100
Eric Dumazet <edumazet@xxxxxxxxxx> wrote:

> > 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()) {  
> 
> A barrier() is needed here, with the current state of affairs.
> 
> IMO, ftrace/x86 experts should take care of this generic issue ?

trace_*_enabled() is a static_branch() (aka. jump label).

It's a nop, where the if block is in the out-of-line code and skipped. When
the tracepoint is enabled, it gets turned into a jump to the if block
(which returns back to this location).

That is, when the tracepoint in the block gets enabled so does the above
branch. Sure, there could be a race between the two being enabled, but I
don't see any issue if there is. But the process to modify the jump labels,
does a bunch of synchronization between the CPUs.

What barrier are you expecting?

-- Steve

> 
> 
> 
> >                 call_trace_sock_send_length(sock->sk, sock->sk->sk_family,
> >                                             sock->sk->sk_protocol, ret, 0);
> >         }
> >         return ret;
> > }
> >
> > What do you think?