Re: [PATCH v3 net-next 2/5] net: tracepoint: replace tcp_set_state tracepoint with inet_sock_set_state tracepoint

[Brendan Gregg]

On Sat, Dec 30, 2017 at 7:06 PM, Yafang Shao <laoar.shao@xxxxxxxxx> wrote:
> On Sun, Dec 31, 2017 at 6:33 AM, Brendan Gregg
> <brendan.d.gregg@xxxxxxxxx> wrote:
>> On Tue, Dec 19, 2017 at 7:12 PM, Yafang Shao <laoar.shao@xxxxxxxxx> wrote:
>>> As sk_state is a common field for struct sock, so the state
>>> transition tracepoint should not be a TCP specific feature.
>>> Currently it traces all AF_INET state transition, so I rename this
>>> tracepoint to inet_sock_set_state tracepoint with some minor changes and move it
>>> into trace/events/sock.h.
>>
>> The tcp:tcp_set_state probe is tcp_set_state(), so it's only going to
>> fire for TCP sessions. It's not broken, and we could add a
>> sctp:sctp_set_state as well. Replacing tcp:tcp_set_state with
>> inet_sk_set_state is feeling like we might be baking too much
>> implementation detail into the tracepoint API.
>>
>> If we must have inet_sk_set_state, then must we also delete tcp:tcp_set_state?
>>
>
> Hi Brendan,
>
> The reason we have to make this change could be got from this mail
> thread, https://patchwork.kernel.org/patch/10099243/ .
>
> The original tcp:tcp_set_state probe doesn't traced all TCP state transitions.
> There're some state transitions in inet_connection_sock.c and
> inet_hashtables.c are missed.
> So we have to place this probe into these two files to fix the issue.
> But as inet_connection_sock.c and inet_hashtables.c are common files
> for all IPv4 protocols, not only for TCP, so it is not proper to place
> a tcp_ function in these two files.
> That's why we decide to rename tcp:tcp_set_state probe to
> sock:inet_sock_set_state.

It kinda feels like we are fixing one exposing-implementation problem
(the missing state changes, which I'm happy to see fixed), by exposing
another (there's no tcp:tcp_set_state because we don't want to put tcp
functions in inet*.c files). Anyway...

If I'm to use sock:inet_sock_set_state for TCP tracing, I'd like
sk->sk_protocol exposed as a tracepoint argument so I can match on
IPPROTO_TCP. Otherwise I'll have to keep digging it out of (void
*)skaddr. (And if we're adding arguments, maybe consider sk_family as
well, to make it easier to see which address arguments to use).

Brendan