Re: [lwip-users] Slow release time of closed TCP PCBs

[David Empson]

"Kieran Mansley" <address@hidden> wrote:
> On Thu, 2009-06-18 at 11:00 +0200, Lou Cypher wrote:
> > Simon Goldsmith <address@hidden> wrote:
> > > What you are seeing is that the PCBs are remaining in a wait-state for
> > > some time. This is to prevent packets from the old (closed) connection
> > > being possibly accept on a new connection. The port stays known to the
> > > stack and it can send RST packets when new packets for that port are
> > > received. However, while these PCBs are not yet freed, they are *not* 
> > > in
> > > an active state any more, thus tcp_slowtmr is correct.
> >
> > Everything is clear, does make sense to me.
> > How long can be "some time"? :)
>
> 2 * TCP_MSL
>
> See tcp_slowtmr() where it handles the "TIME-WAIT" pcbs
>
> For the default TCP_MSL, that would be 2 minutes.

One thing I don't understand about this discussion: the tcp_alloc() function 
tries to allocate a new PCB, but if that fails, it tries killing a PCB in 
the TIME_WAIT state (picking the oldest one), then it retries the 
allocation.

This should mean that even if all PCBs are used, it will be possible to 
start a new connection as long as at least one recent connection has closed 
(and has got far enough through the FIN/ACK handshakes, but hasn't reached 
the 2 * TCP_MSL timeout).

This code is similar in LWIP 1.2.0, 1.3.0 and CVS-head, and it works fine 
for us using the raw API, but I haven't looked at implications for higher 
level APIs.

What piece of information am I missing, and why isn't it working for Lou?