mbox series

[net-next,00/13] net: add more tracepoints to TCP/IP stack

Message ID 20210805185750.4522-1-xiyou.wangcong@gmail.com (mailing list archive)
Headers show
Series net: add more tracepoints to TCP/IP stack | expand

Message

Cong Wang Aug. 5, 2021, 6:57 p.m. UTC
From: Cong Wang <cong.wang@bytedance.com>

This patchset adds 12 more tracepoints to TCP/IP stack, both
IPv4 and IPv6. The goal is to trace skb in different layers
and to measure the latency of each layer.

We only add information we need to each trace point. If any other
information is needed, it is easy to extend without breaking ABI,
see commit 3dd344ea84e1 ("net: tracepoint: exposing sk_family in all
tcp:tracepoints").

And similar to trace_qdisc_enqueue(), we only intend to trace
success cases, because most (if not all) failure cases can be traced
via kfree_skb() even if they are really interesting.

Lastly, per previous discussion, trace ring buffer is only accessible
to privileged users, it is safe to use a real kernel address with %px.

Qitao Xu (13):
  net: introduce a new header file include/trace/events/ip.h
  ipv4: introduce tracepoint trace_ip_queue_xmit()
  tcp: introduce tracepoint trace_tcp_transmit_skb()
  udp: introduce tracepoint trace_udp_send_skb()
  udp: introduce tracepoint trace_udp_v6_send_skb()
  ipv4: introduce tracepoint trace_ip_rcv()
  ipv6: introduce tracepoint trace_ipv6_rcv()
  ipv4: introduce tracepoint trace_ip_local_deliver_finish()
  udp: introduce tracepoint trace_udp_rcv()
  ipv6: introduce tracepoint trace_udpv6_rcv()
  tcp: introduce tracepoint trace_tcp_v4_rcv()
  ipv6: introduce tracepoint trace_tcp_v6_rcv()
  sock: introduce tracepoint trace_sk_data_ready()

 include/trace/events/ip.h   | 140 ++++++++++++++++++++++++++++++++++++
 include/trace/events/sock.h |  19 +++++
 include/trace/events/tcp.h  |  27 ++++++-
 include/trace/events/udp.h  |  74 +++++++++++++++++++
 net/core/net-traces.c       |   7 ++
 net/ipv4/ip_input.c         |   9 ++-
 net/ipv4/ip_output.c        |  10 ++-
 net/ipv4/tcp_input.c        |   8 ++-
 net/ipv4/tcp_ipv4.c         |   2 +
 net/ipv4/tcp_output.c       |   8 ++-
 net/ipv4/udp.c              |  17 ++++-
 net/ipv6/ip6_input.c        |   5 +-
 net/ipv6/tcp_ipv6.c         |   3 +
 net/ipv6/udp.c              |   9 ++-
 14 files changed, 327 insertions(+), 11 deletions(-)
 create mode 100644 include/trace/events/ip.h

Comments

Cong Wang Aug. 6, 2021, 2:22 a.m. UTC | #1
On Thu, Aug 5, 2021 at 11:57 AM Cong Wang <xiyou.wangcong@gmail.com> wrote:
>
> From: Cong Wang <cong.wang@bytedance.com>
>
> This patchset adds 12 more tracepoints to TCP/IP stack, both
> IPv4 and IPv6. The goal is to trace skb in different layers
> and to measure the latency of each layer.

It looks like we should not trace them at the end of each target function,
instead we have to trace them in the beginning, otherwise the ordering
is reversed in the output.

I will send V2 tomorrow.

Thanks.