Re: [PATCH net-next 2/2] tcp: handle TCP_TIME_WAIT/TCP_NEW_SYN_RECV in tcp_set_state tracepoint
From: Eric Dumazet <hidden>
Date: 2017-11-09 14:52:52
Also in:
lkml
On Thu, 2017-11-09 at 14:26 +0000, Yafang Shao wrote:
quoted hunk ↗ jump to hunk
When TCP connetion in TCP_TIME_WAIT or TCP_NEW_SYN_RECV state, it can't get the sport/dport/saddr/daddr from inet_sock. trace_tcp_set_state may be called when the oldstate in these two states. Signed-off-by: Yafang Shao <redacted> --- include/trace/events/tcp.h | 33 ++++++++++++++++++++++++++------- 1 file changed, 26 insertions(+), 7 deletions(-)diff --git a/include/trace/events/tcp.h b/include/trace/events/tcp.h index 07cccca..1982a71 100644 --- a/include/trace/events/tcp.h +++ b/include/trace/events/tcp.h@@ -196,7 +196,6 @@ ), TP_fast_assign( - struct inet_sock *inet = inet_sk(sk); struct in6_addr *pin6; __be32 *p32;@@ -204,14 +203,34 @@ __entry->oldstate = oldstate; __entry->newstate = newstate; - __entry->sport = ntohs(inet->inet_sport); - __entry->dport = ntohs(inet->inet_dport); + if (oldstate == TCP_TIME_WAIT) { + __entry->sport = ntohs(inet_twsk(sk)->tw_sport); + __entry->dport = ntohs(inet_twsk(sk)->tw_dport); - p32 = (__be32 *) __entry->saddr; - *p32 = inet->inet_saddr; + p32 = (__be32 *) __entry->saddr; + *p32 = inet_twsk(sk)->tw_rcv_saddr; - p32 = (__be32 *) __entry->daddr; - *p32 = inet->inet_daddr; + p32 = (__be32 *) __entry->daddr; + *p32 = inet_twsk(sk)->tw_daddr; + } else if (oldstate == TCP_NEW_SYN_RECV) { + __entry->sport = inet_rsk(inet_reqsk(sk))->ir_num; + __entry->dport = ntohs(inet_rsk(inet_reqsk(sk))->ir_rmt_port); + + p32 = (__be32 *) __entry->saddr; + *p32 = inet_rsk(inet_reqsk(sk))->ir_loc_addr; + + p32 = (__be32 *) __entry->daddr; + *p32 = inet_rsk(inet_reqsk(sk))->ir_rmt_addr; + } else { + __entry->sport = ntohs(inet_sk(sk)->inet_sport); + __entry->dport = ntohs(inet_sk(sk)->inet_dport); + + p32 = (__be32 *) __entry->saddr; + *p32 = inet_sk(sk)->inet_saddr; + + p32 = (__be32 *) __entry->daddr; + *p32 = inet_sk(sk)->inet_daddr; + }
Wow. Since all three variants of sockets (full sockets, request sockets, timewait sockets) are all hashed into ehash table these days, they all have the fields at the same offset For IPv4, that would be : __sk_common.skc_daddr (or inet_daddr) __sk_common.skc_rcv_saddr (or inet_rcv_saddr ) __sk_common.skc_dport (or inet_dport) __sk_common.skc_num (or inet_num) Look at __inet_lookup_established() and INET_MATCH() : They deal with the three variants, without having to look at sk_state. If you were using the fields that are common to all sockets, no need to add all this unnecessary complexity.