Re: TCP RESET after SACK in Kernel 2.6

From: Christian Schwarz <redacted>
Date: Thu, 16 Oct 2008 10:36:57 +0200

[ netdev is the place to post if you actually want a
  knowledgable person to look into your problem ]

We're running a Debian Linux router with masquerading enabled
to access a customers site.  The customer connection has a lot of
packet loss, so we receive some TCP SACK-messages.

Our old system with Debian 3.1 and Kernel 2.4.xx handled this correctly,
but after upgrading to Debian 4.0 and Kenrel 2.6.26.5, sometimes the
SACK-queries are answered with a TCP RESET by the kernel.

Without any example tcpdump traces it is impossible for anyone to help
debug this problem.

Background question: we have the following setup:

   client host  ->  router w/ masquerading  ->  target host

When the target host sends a SACK-query, is it the router's job to
resend the packets or is this the client host's job?

The end node TCP stack does this, the router doesn't do this.