Date: Wed, 23 Jan 2002 19:27:45 -0500 From: Joseph Ishac <email@example.com> Subject: Observation: Retransmitted data with FIN bit set. (Possible tcptrace bug) Message-ID: <20020123192745.B2979@sunfire.grc.nasa.gov>
I am currently running several network tests in a closed network
consisting of three machines. The goal is to send a file from one
machine to another through the third machine, which transparently imposes
an artificial delay.
In performing these tests, I've observed a rather strange situation which
causes tcptrace to report the incorrect value in several fields (such as
"ttl stream length" and "missed data").
Analysing the connection revealed that there were two FINs. The first FIN
occurs logically in the final data packet. However, a loss occurs within
the final window of data, and the resulting retransmission also carries a
FIN. Also, the retransmission carries the sequence number of the data
lost, which is lower than the value in the original FIN.
The files attached help give a visualization of the connection that is
causing the problem. Hopefully the files will clear up everything I left
out in my babbling above :).
So, I am unsure what the correct behavior should be...
According to RFC 793, the FIN bit represents 'no more data to send', thus
tacking it onto the retransmission does not seem to be counter intuitive -
since the sender has no more data to send. However, setting the FIN bit
isn't necessary either, since the stream is reliable. I've made a quick
pass through RFC 793 and could not find anything that indicates which
behavior is correct.
In either case, my guess is that the retransmission with the FIN bit set
is 'confusing' tcptrace, causing it to report erroneous values. If this
is the case _and_ the behavior of the connection is valid, then tcptrace
would likely need to use the FIN packet with the highest sequence number
(or just ignore FIN packets whose sequence number is lower than previous
FINs) to calculate the appropriate values. However, I have not had a
chance to look into the tcptrace code to verify either the problem or
justify the solution.
Any ideas/thoughts on what the correct behavior should be?
NASA Glenn Research Center
enc. 4 files
dump.txt : Human readable tcpdump output for the connection of interest.
sequence_graph.xpl : Time sequence graph of the connection. (Xplot)
tcptrace.txt : Broken output of tcptrace.
tcp-reduce.txt : Broken output of tcp-reduce.
PS: A few things I forgot to mention:
- The correct size of the file being transferred is 28960 bytes
- The machines sending and receiving the file are running OpenBSD 2.9
- The bridge imposing the delay is running Solaris 8
- A similar note is being sent out to the folks in charge of tcp-reduce.
This archive was generated by hypermail 2b30 : 01/24/02 EST