[tcpm] meaning of "idle" for TCP Keep-Alives (was: 793bis ready to go?)

[Neal Cardwell <[e-mail]>]

On Fri, Feb 12, 2021 at 10:19 AM Michael Tuexen <
[e-mail]> wrote:

> > On 11. Feb 2021, at 06:22, Neal Cardwell <ncardwell=
> [e-mail]> wrote:
> >
> > The wording in the Congestion Control section looks great to me.
> >
> > > Minor:
> https://tools.ietf.org/html/draft-ietf-tcpm-rfc793bis-20#section-3.7.4
> >
> > I agree with Yuchung that it would be nice to clarify what is meant by
> "idle" connections (in the pre-existing text), with respect to keep-alives.
> It seems at least the historical behavior for two of the major
> implementations out there interpret this in different ways, and this has
> caused some confusion in the developer community about what semantics to
> expect from TCP keep-alives. I guess the rfc793bis-20 draft does have some
> nice new text (that does not seem to be in RFC793 or RFC1122) that seems to
> shed some light on this:
> >
> >    Keep-alive packets MUST only be sent when no sent data is
> >    outstanding, and no data or acknowledgement packets have been
> >    received for the connection within an interval (MUST-26).
> >
> >
> > That sentence helps quite a bit. It seems to correspond to the last 20
> years of Linux TCP keep-alive behavior, but not the BSD behavior (at least
> BSD TCP as documented in Stevens volume 2; I have not checked more recent
> BSD kernels).
> Can elaborate what Linux behaviour you are referring to an what BSD
> behaviour you are referring to?


BSD: My reading of the behavior in older BSD releases (BSD-4_2 [1983]
through BSD-4_4_Lite2 [1995], or Stevens volume 2 [1995]) is that a TCP
endpoint resets the keep-alive timer (tp->t_timer[TCPT_KEEP]) to its full
keep-alive interval upon receiving a packet, and then when the timer fires
it sends a keep-alive probe packet (case TCPT_KEEP in tcp_timers() in
tcp_timer.c). When that keep-alive timer fires it seems to send a
keep-alive packet irrespective of whether the TCP endpoint currently has
data outstanding in the network, or is sending ZWP packets.

Linux: Starting in Linux 2.3.41 in Jan 2000, and continuing through current
releases, the Linux TCP keep-alive timer code in tcp_keepalive_timer() has
two simple extra checks introduced:
  https://elixir.bootlin.com/linux/2.3.41/source/net/ipv4/tcp_timer.c#L729
The checks basically say that if the keep-alive timer goes off but the TCP
endpoint has data outstanding in the network (RTO timer is set), or
outgoing data queued and waiting for a non-zero receive window (ZWP timer
is set), then the code merely resets the keep-alive timer, and
short-circuits all the keep-alive processing that would send a keep-alive
probe packet or check to see if the connection should be torn down. This
behavior is similar to the rfc793bis-20 text,  "Keep-alive packets MUST
only be sent when no sent data is  outstanding", except that it also skips
sending a keep-alive when the sender is in the process of sending repeated
zero window probes (which would also serve to tell whether the remote
endpoint is reachable).

best,
neal