Path: utzoo!utgpu!jarvis.csri.toronto.edu!cs.utexas.edu!tut.cis.ohio-state.edu!ucbvax!OPUS.CRAY.COM!dab
From: dab@OPUS.CRAY.COM (Dave Borman)
Newsgroups: comp.protocols.tcp-ip
Subject: Re:  TCP Retries
Message-ID: <9002161823.AA02135@opus.cray.com>
Date: 16 Feb 90 18:23:41 GMT
Sender: daemon@ucbvax.BERKELEY.EDU
Organization: The Internet
Lines: 46

> I'm testing my port of a version of 4.3BSD TCP/IP code.  In
> doing so, I discard in the loopback driver every Nth packet
> (say, N=10).  This results in the MBUF pool being quickly
> depleted.  I discovered that the TCP input routine will
> save all out of order TCP packets waiting for the TCP retry
> to fill in the gaps.  Hence most of the MBUF pool is
> enqueued on this TCP fragment/reassembly list for this
> TCPCB.  Also, the sender continues to get ACKs advertising
> the largest window since the receive socket buffer is
> empty.
> 
> Nothing particularly evil happens and since this is
> a very degenerate case, perhaps I should just shrug
> my shoulders and go on with life.  But it does bother
> me a bit - rightfully so?  (Couldn't the window advertised
> by the receiver close, or a limit made on the number
> of out-of-order TCP packets on a TCPCB?)
> 
> Thanks for any help.
>   Dave (wiltzius@lll-lcc.llnl.gov)

There are a few things to be aware of.  First, the TCP
reassembly queue will not grow for ever.  It can only
hold as much data as was offered in the window.  So,
in theory, this is no different that having the user
process stop reading the socket, and letting the
receive queue fill up with data.

But in actuallity, there are some differences.  The 4.3
code does not do any compression on the data in the
reassembly queue, like it does on the receive queue for
the socket.  Thus, if you have a window of, say, 4k, and
you have the data coming in in 1 byte packets, and you
loose the first packet, you could wind up having over
four thousand mbufs stuck in the reassembly queue.

There are two things that should be done to improve
this situation:
	1) fix the reassembly code to do compaction.
	2) When you run out of mbufs, flush the tcp
	   reassembly queues.

Both of these features have been in our networking
code for several releases.

		-Dave Borman, dab@cray.com