tcp heisenbug

tcp heisenbug

[Andrew Morton]

I just booted my main workstation into 2.6.9-rc3-mm3 (approx).

fetchmail broke.  It's fetching mail from localhost (ssh port-forwarded to
remote servers) and then for some reason it wants to open an SMTP
connection to localhost.  It was failing, some error message about that
SMTP connection.

While I was poking at it, it started to work.  `uptime' said 6 minutes.  So
I'd be suspecting that there's a jiffy wrap bug in TCP-to-localhost
introduced since 2.6.9-rc3.

I rebooted to do more diagnosis and the damn thing worked fine.

I can't see anything suspicious in the post-rc3 net/ diff.  But if anyone
has fiddled with time-related things in networking, please double-check
your work?

Re: tcp heisenbug

[David S. Miller]

On Fri, 8 Oct 2004 01:41:51 -0700
Andrew Morton <akpm@osdl.org> wrote:

> fetchmail broke.  It's fetching mail from localhost (ssh port-forwarded to
> remote servers) and then for some reason it wants to open an SMTP
> connection to localhost.  It was failing, some error message about that
> SMTP connection.
> 
> While I was poking at it, it started to work.  `uptime' said 6 minutes.  So
> I'd be suspecting that there's a jiffy wrap bug in TCP-to-localhost
> introduced since 2.6.9-rc3.

That's possible.

It could also be something that only happens the first time you try
to talk to a remote system.

You're using e1000 on x86 right?  Maybe force TSO off via ethtool
right after the interface is brought up and see if that makes it
go away?

Re: tcp heisenbug

[Andrew Morton]

"David S. Miller" <davem@davemloft.net> wrote:
>
> On Fri, 8 Oct 2004 01:41:51 -0700
> Andrew Morton <akpm@osdl.org> wrote:
> 
> > fetchmail broke.  It's fetching mail from localhost (ssh port-forwarded to
> > remote servers) and then for some reason it wants to open an SMTP
> > connection to localhost.  It was failing, some error message about that
> > SMTP connection.
> > 
> > While I was poking at it, it started to work.  `uptime' said 6 minutes.  So
> > I'd be suspecting that there's a jiffy wrap bug in TCP-to-localhost
> > introduced since 2.6.9-rc3.
> 
> That's possible.
> 
> It could also be something that only happens the first time you try
> to talk to a remote system.

It doesn't seem like that.  I'd already established three ssh sessions to
three remote servers (zip,  digeo, osdl) and fetchmail was able to contact
those three servers via connection to the three local ssh clients.

It was only when fetchmail tried to talk to sendmail on localhost that
things failed.

> You're using e1000 on x86 right?

Yup.  At 100 mbps

>  Maybe force TSO off via ethtool
> right after the interface is brought up and see if that makes it
> go away?

The problem went away on subsequent reboots :(

But the failure was when communicating with sendmail on localhost.