Re: TCP in TIME_WAIT response to dup FIN

Re: TCP in TIME_WAIT response to dup FIN

[Guo, Liang]

> I'm performing a benchmark test to a local linux website. When
> I monitor packet sequences at the website with tcpdump, I found
> that when tcp socket is in TIME_WAIT state, the default response to a
> retransmitted FIN packets is to silently drop them. Is this the correct
> way of implementing TIME_WAIT? The website was runing on
> 2.2.19 kernel, and when I change it to 2.4.9, tcp becomes more
> active and start sending back FIN-ACKs.
> 
> I saw different explainations on the purpose of TIME_WAIT state,
> some of them (including Stevens' book) say that it can deal with
> losses of the last FIN or FIN-ACK packet in addition to
> preventing reuse of same address-port pair immediately.
> 
> I also glimpsed through TCP implementations in both kernels.
> Although both claim to be conformant to RFC 1122, the code
> themselves are quite different. And I couldn't find 
> anywhere in the 2.2.19 kernel implementation that tells
> TCP to respond to duplicate FINs except in tcp_timewait_state_process:
> 
>         /* Ack old packets if necessary */ 
>         if (!after(TCP_SKB_CB(skb)->end_seq, tw->rcv_nxt) &&
>             (len > (th->doff * 4) || th->fin))
>                 return TCP_TW_ACK; 
>         return 0; 
> 
> Which I suspect contains some bug.
> I've searched the archive and found a related message was posted
> almost 3 years ago:
> http://marc.theaimsgroup.com/?l=linux-kernel&m=93293686728605&w=2
> Seems that the "|| th->fin" condition was the ad-hoc patch for
> the problem. And if it is, shouldn't the "&&" be "||" in the
> if condition?
> 
> Again, my question is whether 2.2.19 is doing correct things, 
> and if not, can I replace tcp implementation in 2.2.19 by
> that in 2.4.9 and how? Or some alternative milder changes?
> 
> Thank you very much.
> 
> 
> Guo, Liang 
> 
> guol@cs.bu.edu                     Dept. of Comp. Sci., Boston Univ.,
> (617)353-5222 (O)                  111 Cummington St., MCS-217,
> (617)375-9206 (H)                  Boston, MA 02215
> 
> 

Re: TCP in TIME_WAIT response to dup FIN

[kuznet]

Hello!

> > that when tcp socket is in TIME_WAIT state, the default response to a
> > retransmitted FIN packets is to silently drop them. Is this the correct
> > way of implementing TIME_WAIT?

No. Linux 2.2 is not expected to have this bug.


> > active and start sending back FIN-ACKs.

No. time-wait never send FINs, just bare ACK.



> >         /* Ack old packets if necessary */ 
> >         if (!after(TCP_SKB_CB(skb)->end_seq, tw->rcv_nxt) &&
> >             (len > (th->doff * 4) || th->fin))
> >                 return TCP_TW_ACK; 
> >         return 0; 
> > 
> > Which I suspect contains some bug.

It should responce right. end_seq == rcv_nxt in the case of retransmitted
FIN, len == th->doff*4 but th->fin == 1. So, it returns TCP_TW_ACK.

So, it looks right, I have no idea what happens in your case.
Apparently, you diagnosed problem wrongly. It would be better
if you sent tcpdump of bad session instead.


> > Again, my question is whether 2.2.19 is doing correct things, 

Supposed to do. If it does not, alas.


> > that in 2.4.9 and how? Or some alternative milder changes?

Milder change is just to install 2.4. :-)

Alexey