Re: [ns] FullTcp & "done"

[Tarik Alj <aljtarik@cholla.inrs-telecom.uquebec.ca>]

Felix,

	good! This is getting fun... I'll be trying your patch. Just to make 
sure we stay SIN :)  could you send me a copy of your tcp-full.h and .cc? 
	
>Date: Sun, 06 Aug 2000 14:42:43 -0400
>From: Felix Hernandez Campos <fhernand@cs.unc.edu>
>X-Accept-Language: en
>MIME-Version: 1.0
>To: Tarik Alj <aljtarik@cholla.inrs-telecom.uquebec.ca>, Haobo Yu 
<haoboy@ISI.EDU>, Limin Wang <lmwang@CS.Princeton.EDU>, "Guo, Liang" 
<guol@cs.bu.edu>
>CC: NS-USERS <ns-users@mash.CS.Berkeley.EDU>
>Subject: Re: [ns] FullTcp & "done"
>
>Well, to make things more clear, I'd like to point out that we are dealing
>with two problems here: invoking the done callback at the right time, and
>reusing FullTcp agents. These two issues are closely related, but they are not
>exactly the same.
>
>Real TCP implementations don't have a done callback, so it is not even well
>defined what we mean by that. Does it mean *both* endpoints are closed? Or
>that the local endpoint is closed and not willing to accept any more packets?
>(I'm not even considering half-closed connections yet...). If we are going to
>use done for reusing agents, it makes sense to invoke done when both endpoints
>are closed. Unfortunately, this is troublesome, since we will have to know
>that both agents are in TCPS_CLOSED before invoking the done callback (in
>fact, this is a bad approximation to any real implementation). This is not
>what we are actually trying to do now, since done is invoked whenever an agent
>considers its connection closed, but it doesn't care about the other endpoint
>(well, by the implemented TCP state machine, the other endpoint is also closed
>or closing the connection). The later definition is enough for webtraf, since
>we can attach the cleanup done callback to the client endpoint (close is
>invoked by the server endpoint, so the client will be the passive closing
>endpoint, and its done callback will be invoked iff both server and client are
>not going to send any more packets.)
>
>I did some significant progress implementing the done callback. First of all,
>I discovered a bug in FullTcp code. tcp-full.h defines a closed_ instance
>variable in FullTcpAgent class, and this is wrong (*). closed_ is already in
>TcpAgent, so it shouldn't be redefined. That was causing a visibility problem
>with TcpAgent::reset(), that did not assign 0 to FullTcpAgent::closed_. After
>removing closed_ from tcp-full.h, we can call finish instead of evalf in
>tcp-full.cc:
>
> // haoboy: Is here the place for done{} of active close?
> // It cannot be put in the switch above because we might need to do
> // send_much() (an ACK)
> // Felix (using Tarik suggestion)
> if (state_ == TCPS_CLOSED) {
>  finish();
> }
>
>We don't need to call cancel_timers() in TCPS_LAST_ACK, since that's now done
>by finish(). I tried this and it works well. I have attached an example tcl
>script. However, I found that this is not enough when the FIN from the passive
>closing endpoint is lost. In this case, the done callback is never invoked.
>The passive closing endpoint keeps retransmitting FIN, with no answer from the
>active closing endpoint. You can test this using my script, if you enable the
>errormodule (it is commented out). The problem here is that FullTcpAgent's TCP
>implementation in NS lacks one state (TIME_WAIT). In real implementations,
>this state is in charge of waiting 2MSL before allowing any socket to reuse
>the port. It is also responsible for acknowledging FIN retransmissions (see
>RFC 793). Since we don't have this state, our FullTcp agent goes directly to
>TCPS_CLOSED. All retransmitted FINs match the first case in recv():
>
> if (state_ == TCPS_CLOSED)
>  goto drop;
>
>This statement is also in the BSD code, so it's correct, and we have to keep
>it here. In BSD's tcp_input(), we have two cases before this one, that will
>send a RST packet for unknown connections or closing connections (using the
>PCB and TCPCB). We don't have these cases in FullTcpAgent (neither we have PCB
>or TCPCB). We don't even have RST packets!
>
>We could hack the code and add a case just before that if statement:
>
> // Felix: bug fix
> // acknowledge FIN from passive closer even in TCPS_CLOSED state
> // (since we lack TIME_WAIT state and RST packets,
> // the loss of the FIN packet from the passive closer
> // will make that endpoint to retransmit the FIN forever)
> if ( (state_ == TCPS_CLOSED) && (tiflags & TH_FIN) )
>  goto dropafterack;
>
>This code solves the FIN problem. My script works fine now. I also tried
>test-all-full and all tests succeeded. However, this is only intended as a
>temporary solution. I feel we should fix FullTcp and add a TIME_WAIT state. I
>guess KF didn't include this state thinking that we can do better in a
>simulator that in real life, but it makes things more difficult to understand
>and creates new situations, increasing the chance of bugs. Additionally,
>adding TIME_WAIT will probably solve the problem with on-the-wire packets,
>since this state will force the agent to wait 2MSL before moving to
>TCPS_CLOSED.
>
>Anyway, adding a TIME_WAIT state is a "major" modification in FullTcp, so I
>will appreciate your feedback.
>
>Cheers,
>
>-- Felix Hernandez
>
>(*) It was added in revision 1.17 for .h and 1.32 for .cc, actually to support
>a done callback through finish().

Tarik