SENDFILE fix not working, on to tcp_cork (Re: HAVE_SENDFILE change not effective (Re: [distcc] Problems with distcc hanging on large compiles))

[Martin Pool]

On 30 Aug 2002, "Hien D. Ngo" <hien at moses.xp.com> wrote:
> 
> Sorry, didn't realize that you wanted the info on the FIN_WAIT1 hang ups still.  The 
> bug seemed easy to reproduce, so I'll rerun with tcp_cork back on when I get a 
> chance.  RH 6 machines seemed to bomb out very quickly.
> 
> So far, so good, though after turning off tcp_cork.  Both test compile loops are 
> running fine on RH 7.2/7.3 and RH 6 test machines.  I'll leave them running for a 
> while to be sure.
> 
> I'll be out of town for the weekend, but if I get a chance, I'll reproduce the 
> FIN_WAIT1 problems, I'll send the data.

I'm wondering whether there is a kernel bug related to TCP_CORK.  I
think I'm using it in an uncontroversial way, but it's a new feature
and there might be something wrong.  Getting `netstat -to` output on
both ends, plus (if possible) a tcpdump may show that the problem is
in the kernel, if it indicates that the tcp stack has got to a state
it should never enter.  One example would be being in FIN_WAIT1 with
data pending and no timers registered.  

Thanks for running all the experiments.

Have a good weekend.

--
Martin