Re: [PATCH] net/tcp: introduce TRACE_EVENT for TCP/IPv4 state transition

From: Song Liu
Date: Thu Nov 09 2017 - 18:42:13 EST



> On Nov 9, 2017, at 10:34 AM, Song Liu <songliubraving@xxxxxx> wrote:
>
>>
>> On Nov 9, 2017, at 10:18 AM, Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
>>
>> On Thu, 9 Nov 2017 15:43:29 +0900
>> Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> wrote:
>>
>>>> +TRACE_EVENT(tcp_set_state,
>>>> + TP_PROTO(struct sock *sk, int oldstate, int newstate),
>>>> + TP_ARGS(sk, oldstate, newstate),
>>>> +
>>>> + TP_STRUCT__entry(
>>>> + __field(__be32, dst)
>>>> + __field(__be32, src)
>>>> + __field(__u16, dport)
>>>> + __field(__u16, sport)
>>>> + __field(int, oldstate)
>>>> + __field(int, newstate)
>>>> + ),
>>>> +
>>>> + TP_fast_assign(
>>>> + if (oldstate == TCP_TIME_WAIT) {
>>>> + __entry->dst = inet_twsk(sk)->tw_daddr;
>>>> + __entry->src = inet_twsk(sk)->tw_rcv_saddr;
>>>> + __entry->dport = ntohs(inet_twsk(sk)->tw_dport);
>>>> + __entry->sport = ntohs(inet_twsk(sk)->tw_sport);
>>>> + } else if (oldstate == TCP_NEW_SYN_RECV) {
>>>> + __entry->dst = inet_rsk(inet_reqsk(sk))->ir_rmt_addr;
>>>> + __entry->src = inet_rsk(inet_reqsk(sk))->ir_loc_addr;
>>>> + __entry->dport =
>>>> + ntohs(inet_rsk(inet_reqsk(sk))->ir_rmt_port);
>>>> + __entry->sport = inet_rsk(inet_reqsk(sk))->ir_num;
>>>> + } else {
>>>> + __entry->dst = inet_sk(sk)->inet_daddr;
>>>> + __entry->src = inet_sk(sk)->inet_rcv_saddr;
>>>> + __entry->dport = ntohs(inet_sk(sk)->inet_dport);
>>>> + __entry->sport = ntohs(inet_sk(sk)->inet_sport);
>>>> + }
>>>> +
>>>> + __entry->oldstate = oldstate;
>>>> + __entry->newstate = newstate;
>>>> + ),
>>>> +
>>>> + TP_printk("%08X:%04X %08X:%04X, %02x %02x",
>>>> + __entry->src, __entry->sport, __entry->dst, __entry->dport,
>>>> + __entry->oldstate, __entry->newstate)
>>>
>>> direct %x of state is not allowed.
>>> This has to use show_tcp_state_name() like it's done in trace_tcp_set_state
>>
>> Hmm, I need to look at trace_tcp_set_state. I'm guessing it is in
>> net-next too?
>
> Yes, in net-next. There are 6 tracepoints under tcp group:
>
> tcp_destroy_sock
> tcp_receive_reset
> tcp_retransmit_skb
> tcp_retransmit_synack
> tcp_send_reset
> tcp_set_state
>
> They are all added recently.
>
>>
>>>
>>> Also I'm missing the reason to introduce another tracepoint
>>> that looks just like trace_tcp_set_state.
>>
>> I want to make sure that perf and trace-cmd can parse the TP_printk(),
>> if it is having helper functions like that in the TP_printk() part,
>> then the libtraceevent needs to be aware of that.
>>
>
> tcp_set_state uses __print_symbolic to show state in text format. I found
> trace-cmd cannot parse that part:
>
> [011] 147338.660560: tcp_set_state: sport=16262 dport=48346 \
> saddr=127.0.0.6 daddr=127.0.0.6 saddrv6=2401:db00:30:317e:face:0:1f:0 \
> daddrv6=2401:db00:30:31e5:face:0:7f:0 oldstate= newstate=
>
> Other parts of the output looks good to me.
>
> Thanks,
> Song

I am not sure whether this is the best approach, but the following patch
fixes the output of perf:

0.44% sport=16262 dport=39362 saddr=127.0.0.6 daddr=127.0.0.6 \
saddrv6=2401:db00:30:317e:face:0:1f:0 daddrv6=2401:db00:30:31e5:face:0:7f:0 \
oldstate=TCP_CLOSE_WAIT newstate=TCP_LAST_ACK

Thanks,
Song