TCP Free-BSD setup behaviour.

[Louis Mamakos]

On Jun 10, 2009, at 9:47 AM, saravana perumal wrote:

>  Hi ,
>
>   Have some behaviour change  with FREEBSD  compared to  LINUX .

You probably ought to verify the behavior against the protocol  
specifications, and not what some other TCP implementation happens to  
to.
>
> 1. When sending the Same  TCP packet once again [ Retranmission of  
> TCP packet ] Whether the Same Identification field [ IP packet]used  
> or not .
> but when seeing that thru packet capture, Free BSD sending the  
> differnt one [ increases sequentially IP Identification]

The IPID header field is used for reassembly of IP fragments, and is  
not of consequence to TCP.  If the protocol stack absolutely knows  
that a TCP retransmission is identical to the previous segment, then  
in theory, it could use the same IPID fields to increase the chance  
that a previously fragmented TCP segment with a lost segment could get  
reassembled with fragments from the retransmission.  Since there's  
much work done to avoid fragmentation in the first place (e.g., using  
Path MTU discovery), this "feature" probably never gets used.

This behavior makes more sense if the TCP implementation keeps around  
a retransmit queue of previously sent packets, rather than simply  
generating new TCP segments with whatever data happens to be in the  
TCP send window at the time of the retransmission attempt.

>
> 2.Retranmission Time is not increasing Linearly with Respect to BSD.  
> not keeping more time interval . AS per RFC
> expecting Retransmission timeout should  increase Linearly. Issue is  
> not seen with Linux Setup

Retransmission time is highly dependent on many factors, like the  
implementation of TCP slow-start, what the TCP stack has estimated as  
the round-trip time, etc.  In the general case, over the span of many  
retransmissions, the sending TCP stack should back-off the retransmit  
times.

>
> 3. Active SYN open state in FREE BSD setup , Does not reach Syn- 
> received State. When Sending syn packet to DUT but  for that FreeBSD  
> is not sending back
> SYN/ACK .  Issue is with Free BSD Setup.Linux works fine,

If the FreeBSD system is doing a TCP active open (e.g., a connect()  
system call), then it sends a SYN to the remote TCP, and expects a SYN/ 
ACK back from the remote system.  At that point, since the remote has  
ACK'd the SYN, it should correctly respond with simply an ACK of the  
remote TCP's SYN, and perhaps any data that might have been  
piggybacked in the ACK segment.  There's no need to retransmit the SYN.

Or is the remote system doing the active open to the FreeBSD system?   
It's hard to believe that the FreeBSD TCP can't respond to an incoming  
TCP segment to a listening socket carrying a SYN segment?

>
> 4.When checking the State - TIME-WAIT
>   Sending FIN and expecting the ACK ;Getting the ACK properly.
>   Sending Data Segment and Expecting the RST signal not getting the  
> RST ; Instead DUT is sending the Last TCP packet.
>
> Issue seen only in Free BSD.

I may be misunderstanding.  When in TIME-WAIT state, the local TCP is  
waiting for a bit in case the "final" ACK of the remote TCP's FIN got  
lost, and the remote retransmits the FIN (and perhaps any data that  
might have been in the window along with the FIN).  The local TCP  
shouldn't generate a RST assuming that the remote's retransmitted TCP  
segment is still within the window.  I'm not sure what's in the "Last  
TCP packet."

>
> Same issue in FIN-WAIT2  & FIN-WAIT1 State also .
>   Sending FIN and Expect the ACK : Getting the ACK
>   Sending Data segment with Data : Expecting the RST signal from  
> DUT ; but got Last transmitted TCP packet[ FIN -ACK]
>

Ditto.

> Any idea about the above scenario would be helpful
>
> Thanks,
> Saravanan.

The TCP in Linux is hardly the reference implementation; with the TCP  
stack in various 4.xBSD varients preceeding it by many years, not to  
mention many others implementations in other platforms.  I'm not sure  
looking for variances between some random Linux TCP stack is really  
the right way to approach testing.

louie